The Second Ops Team

On June 30, 2026, Anthropic released Claude Sonnet 5. On the benchmark suite Anthropic used to justify calling it "the most agentic Sonnet yet," the model scored 81.2 percent on OSWorld-Verified. It scored 80.4 percent on Terminal-Bench 2.1. It reads screens, clicks buttons, types commands, drafts patches, and files pull requests. It is a synthetic worker with real competence on real tools.

One week later, on July 7, 2026, Forbes ran a piece under the headline "Why 40% Of Agentic AI Projects May Be Canceled By 2027." The number came from a Gartner press release dated June 25, 2025, which Gartner has stood behind and repeated through 2026. Anushree Verma, the Gartner analyst on the record, listed three causes: escalating costs, unclear business value, and inadequate risk controls.

Look at the two data points together and one thing jumps out. The models are competent. The projects still cancel. The failure has moved out of the model and into the operating structure around it. And the operations problem is the one nobody is putting on the budget line.

Every AI agent an operations team deploys creates a second operations team that has to run the first one. That second team has payroll, tools, on-call rotations, dashboards, and a monthly bill. Companies that do not staff it end up in Gartner's forty percent. Companies that do end up in the sixty percent that ship. This is the honest picture of AI automation strategy for operations teams in the second half of 2026.

The Pilot That Never Grew Up

Zen van Riel's March 2026 field survey of 650 enterprise technology leaders found that 78 percent of enterprises had at least one AI agent pilot in flight. Fourteen percent had scaled one to organization-wide operational use. The rest were stuck. Not stuck on the model. Not stuck on the API. Stuck on the operating structure around the agent.

Read the survey commentary and the same word keeps appearing. Ownership. The teams that built the pilot were data science teams, applied AI teams, a business unit sponsor, and a time-boxed budget. That is the structure that gets a demo to work. It is not the structure that keeps an agent working at 3 a.m. on a Sunday six months later, when the upstream schema changes and the tool fails silently.

Pilots are projects. Production is a job. Somewhere between those two facts, the money leaks out.

Deloitte's 2026 Tech Trends report put the same idea in the language of the CIO. Agentic strategy is not a technology decision. It is an organizational design decision. Which is a polite way of saying the same thing the survey said. You need a team. That team is not the pilot team. And you have not hired it.

What The Second Ops Team Actually Does

Once an agent leaves the demo, five kinds of work start showing up every week. None of the five is glamorous. Skip any of them and the agent stops shipping value.

The first is evaluation. Every agent needs a running harness of test cases that reflect real production traffic, not the golden path the pilot used to close its budget. When a model rolls, when a tool changes, when a runbook is updated, the harness runs. The harness lives forever. Someone owns it, updates it, and reads its output. That someone is on payroll.

The second is drift monitoring. Agents fail slowly. The failure mode is not a crash. It is a small increase in the rate at which the agent gets the wrong answer. In a customer service queue, an extra two percent of misrouted tickets shows up as a slow rise in escalation volume six weeks after deployment. Someone has to watch for that curve. In practice, that someone reads dashboards and writes weekly reports the same way an FP&A analyst reads margin reports. The tools are different. The discipline is identical.

The third is cost accounting. Token bills scale with usage in ways that no CFO in 2024 modeled. In 2026 the cost of running an agent to close one ticket can be five cents or fifty dollars, depending on how many tools it calls, how long it thinks, and how many times it retries. A ticket queue with ten thousand tickets a week is a real line item. Someone has to tie every agent action back to a cost center and every cost center back to a service level.

The fourth is incident response. When the agent files a bad pull request, escalates the wrong customer, or approves a refund it should have denied, someone answers for it. There is a page. There is a runbook. There is a rollback. There is a postmortem the next day. PagerDuty's SRE Agent, updated in its Spring 2026 release, participates in incident response inside Slack and Microsoft Teams. It does not remove the need for a human on-call. It changes the shape of what that human does.

The fifth is scope expansion review. Every business unit sponsor wants the agent to do a little more. A customer success VP wants it to draft renewal emails. A finance director wants it to reconcile invoices. Every expansion is a new integration, a new failure mode, a new eval set, a new page target. Someone has to say no. Or say yes and staff the addition. That someone is a manager, not an engineer.

The second ops team does those five things. It is a real team. The second ops team is the AI budget. Anything else is optimism dressed as a plan.

The Auditability Wall

Before the agent can act on a system, the system has to be legible to the auditor. That auditor is now a human plus an evaluation harness. The most common blocker in enterprise AI agent integration is that the internal systems the agent has to touch were never designed to be audited at the action level.

An order management system logs orders. It does not log the reason a field was updated. A CRM logs a call. It does not log the intent inference the caller expressed. A ticketing system logs a resolution. It does not log which knowledge base article the agent used to draft it. When the agent starts acting, all of those actions need traceability so the eval harness can score them and the incident responder can reconstruct them.

That audit work is engineering work. It is invisible to the pilot. It is discovered on the way to production. It is measured in weeks per integration. The March 2026 survey put integration blockers at 95 percent of IT leaders. That is not a data quality problem. It is an observability problem. Every system the agent touches has to be instrumented before the agent can touch it usefully.

Companies that build the audit surface first ship agents that work in production. Companies that skip it ship pilots that never leave the demo. The order is the strategy.

An AI Automation Strategy For Operations Teams That Ships

Every good AI automation strategy for operations teams in 2026 fits four moves. The order matters.

Move one. Capture the process knowledge that lives in your best operator's head. Interview the senior on-call engineer, the tenured support agent, the seasoned buyer. Write down the branching logic they run automatically. This is the raw material. It is also the audit specification. If they cannot articulate the decision tree, the agent cannot execute it and no reviewer can score it.

Move two. Instrument every system that specification touches. Log actions with intent, not just outcomes. Route those logs to one queryable surface. This is the auditability layer. The AWS Well-Architected Generative AI Lens, updated in 2026, calls this the observability boundary. It is where every future evaluation, every future incident, and every future scope review will resolve.

Move three. Wire the agent to a narrow, well-scoped part of that surface. Let it act on a defined class of task. Do not let it wander. Constrain its tool list. Constrain its allowed workflows. Constrain its rollback path. The Gartner analysts who wrote the forty percent number named "unclear business value" as the cause of most cancellations. Unclear value comes from unbounded scope. Bounded scope is measurable.

Move four. Staff the second ops team. Give it a budget line. Give it an org home, most often inside the operations organization the agent serves, occasionally inside a shared platform function. Give it a hiring plan that includes an evaluation engineer, a drift analyst, a cost accountant, and a manager. Give it on-call responsibility. Give it a quarterly business review with the executive sponsor. Treat it like the payroll it is.

That is the whole strategy. It fits on an index card. Almost nobody executes it in that order.

Where The Money Actually Leaks

Gartner's three failure causes read like independent problems. They are one problem in three costumes.

Escalating costs come from unmetered usage. A pilot on ten tickets a day costs eight dollars. The same agent on ten thousand tickets a day becomes a large monthly line item on a spreadsheet no one built. The second ops team catches this in week two. Without a second ops team, the finance director catches it in month six and cancels the project.

Unclear business value comes from unbounded scope. A pilot that solved one workflow well gets asked to solve five. Each addition dilutes the KPI. The pilot's original ROI becomes untraceable. The second ops team says no or attaches metrics to every yes. Without a second ops team, the business unit sponsor loses the thread and the executive committee kills the program.

Inadequate risk controls come from missing auditability. When the agent takes a bad action, the postmortem cannot reconstruct why. Trust collapses. Legal gets involved. The agent goes into a review that ends in a shelf. The second ops team owns the audit surface and the review protocol. Without a second ops team, the first bad action becomes the last action.

Three failure causes. One missing layer. The layer that runs the agent after the pilot ends.

The Vendor Will Not Save You

PagerDuty shipped its SRE Agent. Datadog shipped Bits AI SRE. AWS shipped a DevOps agent that one case study credits with a 77 percent MTTR reduction on a customer's incident response workload. Every observability vendor now has an agent in the box. Every integration vendor now has an agent in the box. Every major ERP vendor has agents on the roadmap and half of them shipped in Q2 2026.

The vendor sells you the agent. The vendor does not sell you the second ops team.

The agent inside the vendor's product has the same operational demands as the agent you build in house. It needs an eval harness that reflects your traffic, not the vendor's. It needs drift monitoring against your KPIs. It needs cost accounting that ties back to your P and L. It needs an incident responder from your team when it does something wrong on your production surface. It needs a scope review process that reflects your policy, not the vendor's default.

The vendor's engineer will help you configure the agent. The vendor's engineer will not sit your standup. Buying the tool is the small decision. Running it is the operating structure decision.

The company that buys the PagerDuty SRE Agent without staffing the second ops team gets exactly the outcome Gartner described. Cost surprises. Value drift. Trust collapse. Cancellation by end of 2027. The company that buys the same agent and staffs the second ops team gets a reliability program that runs quieter every quarter.

Same product. Two outcomes. The difference is a payroll line the CFO did not know to write.

AMD Staffed The Ops Layer

AMD's HR agent case, cited in most 2026 enterprise AI reports, delivered an 80 percent reduction in HR inquiry resolution time and a 70 percent employee satisfaction score within the first ninety days. Every vendor slide deck now cites it. Almost every citation gets the lesson wrong.

Look past the model. Read the AMD case and the operating structure jumps out. AMD's team scoped the agent narrowly. They built an evaluation harness before launch. They routed exceptions to a human queue with a service level agreement. They tracked cost per inquiry from day one. They ran a weekly review with the HR leadership team. AMD staffed a second ops team for the HR agent. That is what the 80 percent reduction is measuring. The model got the case in the door. The second ops team is what kept it there.

Any company can license the same model AMD used. Almost no company copies the operating structure. That is why the case is famous and rare at the same time.

Where To Start This Quarter

Pick one operations workflow. Not five. One. Choose one where the branching logic is clear, the current cost is measurable, and the failure modes are recoverable. Good starting points include refund approval below a threshold, inbound support triage, tier one HR inquiries, weekly reconciliation of two ledgers, and alert acknowledgment on a defined class of incidents.

For that one workflow, write the audit specification first. What action, by whom, with what input, against what rule, with what outcome. Store it. Instrument the systems that support it so every action lands on that specification. Only then wire the agent.

Give the agent the smallest useful tool list. Give it a hard rollback path. Give it a defined rate limit that a finance director can put in the budget. Assign a named engineer to own its evaluation harness. Assign a named analyst to own its drift dashboard. Assign a named manager to own its scope reviews. Give the workflow's business owner a quarterly review with those three people.

Ship it. Measure it. Grow the second ops team only after the first workflow proves the model. This is how the sixty percent that never get canceled actually operate.

The Verdict

The mistake is treating AI as a purchase. Agent deployment is closer to a hire than to a piece of software. Hires need managers, reviews, tools, on-call coverage, and payroll. Every honest AI automation strategy for operations teams in 2026 begins with the second ops team on the budget. Every dishonest one begins with a model name on a slide.

The model is a commodity by July 2026. Claude Sonnet 5, GPT-5 Codex, Gemini 2.5, and open weights within twenty points of them are all available at prices that fall every quarter. The differentiator has moved. It sits inside the operating structure that runs the agent. Companies that architect that structure win the next five years. Companies that shop for tools spend the next five years explaining cancellations to their boards.

Architecting is the word. This is a design decision that touches org charts, budget lines, and audit surfaces. It requires someone who has seen the second ops team pattern work at scale, and knows where the money leaks when the pattern is missing.

Agor AI Advisory designs the second ops team from day one. We start with the workflow that will pay for itself first. We write the audit specification. We stand up the evaluation harness. We name the roles. We build the review cadence. We hand it to your operations organization as a running program, not a slide deck. We stay involved through the first ninety days of production because those are the days where every future cost surprise, value drift, and trust collapse gets prevented.

The forty percent Gartner expects to cancel their agentic projects by 2027 are the companies that will still be shopping for tools next quarter. The sixty percent that ship are already thinking in operating structures. Which side of that line your operations team lands on will be decided this quarter, not next year.