AI Strategy
The 2027 AI team: who to hire, what to hedge, and how to survive an outage
The labs that sell enterprise AI hedge their own supply across clouds, chips and contract terms, while most of their customers plan around one vendor. What November's Microsoft 365 outage showed, and the four roles a 2027 AI team needs.
At 16:17 UTC on Thursday 12 November, Exchange Online began refusing mailbox connections. Within hours the incident covered Teams, SharePoint, Microsoft 365 Copilot and Defender XDR, and broad recovery took roughly 24 hours.
Nobody attacked Microsoft, and no AI agent went rogue. That is why the outage belongs in a 2027 plan. It showed how much enterprise AI now rests on a single platform, and how few people can still run the work by hand when that platform stops.
This is the fourth article in our 2027 series. The readiness checklist ranked what to fix, the budget guide sized the spend, and the DPDP article mapped the regulatory exposure. This one covers who owns it all, and what happens when it breaks. It expands priorities 7 and 8 of the checklist: vendor portability, and people and operating model.
Questions this guide answers
What did the November 2026 Microsoft 365 outage mean for enterprise AI?
Microsoft said the cause was an issue during a network capacity upgrade, not a cyberattack or an AI agent. The lesson is indirect: Copilot and every workflow built on the same tenant stopped with the mailbox. AI inherits the availability of the platform under it. See the outage nobody attacked.
How do frontier AI labs plan capacity, and what can enterprises copy?
They spread supply across several clouds and chip families, keep part of their commitments cancellable, and expect models to turn over within months. Enterprises can copy three habits: a tested second path, contract terms matched to model life, and a deterministic fallback. See how to plan like a frontier lab.
Who should be on an enterprise AI team in 2027?
Four roles: an AI platform and reliability engineer, a context and knowledge architect, an AI governance and eval lead, and a product owner embedded in the business line. The governance lead should report to risk, not to the head of AI. See the four roles.
What should the board track each quarter?
One number per quarter: owned and mapped workflows in Q1, tested second paths in Q2, restore time in a failover drill in Q3, and measured P&L impact in Q4. See the roadmap.
1. The outage nobody attacked
The cause matters, so state it precisely. Microsoft said the incident was a technical issue “not related to cyber or AI / agents”. It traced the problem to an earlier network capacity upgrade, exposed by a later network change. A full post-incident report is still to come.
Two lessons sit inside the timeline.
Single-vendor fragility. Copilot went down with the mailbox. Any agent, workflow or approval chain that reads mail, calendars or SharePoint through that one tenant stopped with it. Most 2027 roadmaps quietly assume that availability is 100%.
Comprehension debt. This is the gap between what your systems do and what your people can still explain and operate by hand. During the outage, automated health checks pulled servers out of service, which pushed load onto the rest. Restarted servers failed again under live traffic. Microsoft first reported no correlated change, then began reverting a network configuration change hours later.
If the operator with the deepest knowledge of the stack needed a day to recover, ask how long your team would need. Ask it about the agent workflows you bought last quarter that nobody on staff has traced end to end. Every abstraction adopted without an owner adds to that debt, and outages are when it falls due.
One detail is worth copying. Microsoft paused all deployments and configuration changes while it investigated. A freeze like that works only if someone has the authority to call it, which is the subject of our piece on change freezes for AI agents.
2. Plan like a frontier lab
Most enterprise AI plans are a licence count and a hope. The labs selling those licences plan for supply risk first, and their own disclosures show how.
Anthropic began its IPO roadshow on 10 November. Its amended prospectus lists about $900.9 billion of committed spend, almost all of it compute and infrastructure, with $451.6 billion falling before the end of 2030. An earlier draft, reported in September, put the figure at $518 billion. The structure matters more than the size.
- Spread. Commitments run across Google, Amazon, Microsoft, SpaceXAI, Broadcom, Nscale, Lambda and AMD, on four chip families: AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs.
- Optionality. About $186.2 billion, roughly a fifth, can be terminated or depends on conditions.
- Short model life. Model versions released in the prior six months produced over 80% of third-quarter revenue.
OpenAI shows the other side: capacity runs out even for the supplier. GPT-6 Astra launched on 3 September. Within days OpenAI paused new sign-ups to its $200 Pro tier because demand strained compute. Its response is breadth: agreements for at least 10 GW of Nvidia systems, 6 GW of AMD and 10 GW of Broadcom accelerators, plus CoreWeave and Stargate beyond Azure.
Neither lab is a model of thrift, and no enterprise should copy the scale. Copy the three habits.
| Planning question | Typical enterprise plan | Frontier lab practice | Rule for 2027 |
|---|---|---|---|
| Where does capability come from? | One cloud, one model family | Several clouds, four chip families | 1. Keep a second path. Every tier-1 AI workflow runs on a second model and cloud, tested each quarter. |
| How long do we commit? | Three-year licence bundle | Long terms for power and silicon, with a fifth cancellable; models turn over in months | 2. Match contract length to model life. Commit long on data, context and governance. Keep model-specific terms to 12 months or less, with swap rights. |
| What happens when supply fails? | Assumed not to | Rations its own flagship, then adds suppliers | 3. Build a deterministic fallback. Each agent workflow has a rules-based or human-queue mode and a runbook staff have rehearsed. |
Rule 3 is the one the Microsoft outage tested. A fallback that depends on the same tenant, identity provider or network as the primary is not a fallback.
3. The four roles a 2027 AI team needs
Pilots stall in production because nobody owns the three things the outage exposed: routing, context and evidence. The “prompt engineer” title and the stand-alone centre of excellence own none of them. Four roles do.
| Role | Owns | Retires | Proof it is working |
|---|---|---|---|
| AI Platform & Reliability Engineer | Multi-model routing, fallbacks, failover drills, cost per task | Each team wiring its own API keys | A tier-1 workflow survives a forced provider switch |
| Context & Knowledge Architect | The rules, exceptions and policy documents agents read; data lineage and access | Prompt engineering as a job | Agent answers trace to a named, current source |
| AI Governance & Eval Lead | Drift testing, DPDP 2027 evidence, audit trails, spend against outcome | Annual model-risk paperwork | A regulator query is answered from logs in days, not weeks |
| Domain-Embedded Product Owner | One workflow, one P&L line, the human runbook | Innovation-lab demos | A measured change in cost, cycle time or loss rate |
The knowledge architect owns what we have called the context layer. The reliability engineer and the eval lead run the control layer between them.
Organise the four as a hub and spokes, with one deliberate split.
- Hub. Platform engineers and knowledge architects sit in one central team. They build routing, context and evaluation once, for every business unit.
- Spokes. Product owners sit inside the business line and report to it. They carry the P&L target, not the platform team.
- Independent line. The governance and eval lead reports to risk or compliance, not to the head of AI. The person who certifies a model should not be measured on shipping it.
A mid-sized firm does not need a large function to start. One named owner per role, even part-time, is better than a 20-person centre of excellence with no accountability for uptime.
This structure is also how comprehension debt gets paid down. The product owner keeps the manual runbook alive. The reliability engineer rehearses the switch. The knowledge architect can say what the agent was told and why.
4. The 2027 roadmap, quarter by quarter
Each quarter pairs one team milestone with one resilience deliverable and one number the board can check. The milestones and targets are our working view, not a sourced benchmark.
| Quarter | Team milestone | Resilience and architecture deliverable | Board KPI |
|---|---|---|---|
| Q1 2027 | Name an owner for each of the four roles. Move governance reporting to risk. | Dependency map of every AI workflow: model, cloud, identity, data source. Tier them by business impact. | 100% of tier-1 workflows have a named owner and a mapped dependency chain |
| Q2 2027 | Platform hub live. First product owners embedded in two business lines. | Model gateway in production. Second model and cloud contracted for tier-1 workflows. | Share of tier-1 workflows with a tested second path |
| Q3 2027 | Eval lead runs monthly drift and DPDP evidence reviews. | First unannounced failover drill. Deterministic fallback and manual runbook for each tier-1 workflow. | Time to restore a tier-1 workflow in the drill, against target |
| Q4 2027 | Product owners report outcomes against the 2027 budget case. | Contract review: model-specific terms at 12 months or less, swap rights in place. | Measured P&L impact per workflow, and share of AI spend on terms under 12 months |
If only one row gets funded, fund Q3. A failover that has never been rehearsed is an assumption, and November showed what assumptions about uptime are worth.
The fourth pillar
The first three articles in this series are only as good as this one. A ranked priority with no owner slips. A budget with no product owner cannot prove it paid. A DPDP control with no eval lead has no evidence behind it.
Microsoft’s outage was not an AI failure, and that is the point. The enterprises that keep working through the next one will have a second path, a rehearsed fallback and people who still understand the process underneath the tooling. Put names against those three things before the 2027 budget closes.
To test your own second path and fallback, see our AI Production Readiness & Governance Review, or request a strategy conversation.
5–7 day review
Find out which of your AI use cases you could not defend to a regulator today.
Filed under