Part 1 of a 3-part series. Part 2, "What to Learn Now," covers the skills that will matter most in 2027. Part 3, "What Shouldn't Surprise You," covers the failures already visible on the shelf.
Prediction season is almost here, and most of what you will read about AI agents in 2027 will come from analysts and venture capitalists. Fair enough. But most of them have never had an agent touch a real ERP. I have. So here is my version, from the operator's seat — starting with what to prepare for.
1. The 40% cull
Gartner forecasts that over 40% of agentic AI projects will be canceled by the end of 2027, driven by escalating costs, unclear business value, and inadequate risk controls. Read that again. Not failed — canceled. Someone with budget authority will look at the pilot, ask what it returned, and kill it.
A second Gartner forecast lands in the same place from the other direction: by 2027, 40% of enterprises will demote or decommission an autonomous agent because of governance gaps identified only after production incidents occur. Note the sequence. The gap surfaces first, usually in an incident. The agent gets pulled second.
The era of the funded experiment is ending. 2027 is the "prove it" phase. Pilots either show return or get cut.
Customer support is where the rollback has already happened, and Klarna is the worked example. Its AI assistant took on work the company compared to roughly 700 full-time agents, and its CEO credits AI with helping shrink Klarna's overall workforce by about 40%. Then, in 2025, Klarna started hiring humans back for the complex cases after service quality slipped. Both halves of that story are true at once: the assistant still handles the routine volume, and the company decided people should handle the moments that matter. Gartner saw the pattern coming — it predicts that by 2027, half of the organizations that planned major AI-driven workforce cuts will abandon those plans. Its survey work backs the skepticism: in late 2025, only 20% of customer service leaders said AI had actually reduced their agent staffing, while 55% reported stable staffing handling higher volumes.
The thread running through all three forecasts is worth naming. Agents rarely get pulled because the technology cannot do the work. They get pulled because a gap — in cost, in value, in control — surfaces late, after money or trust has already been spent. Governance doesn't prevent the bill from coming due. It hands it to you earlier, when it's cheaper.
One honest caveat before the headlines write themselves: whole jobs rarely disappear overnight. Tasks get absorbed, and roles reshape around what's left. The Iceberg Index, built by MIT and Oak Ridge National Laboratory, puts current AI's technical reach at about 11.7% of the US labor market — roughly $1.2 trillion in wages — and that measures task exposure, not pink slips. McKinsey's count has the same shape: 57% of US work hours are technically automatable, but more than 70% of the skills employers want show up in both automatable and non-automatable work. The skills don't disappear. How they're applied changes. Plan for the reshaping, not the disappearance.
2. Agents graduate from copilots to actors
The agents getting the attention right now are chat copilots. The agents that matter in 2027 are actors: systems that write to your ERP and CRM, issue purchase orders, change permissions, post journal entries.
Gartner puts numbers on the shift. It expects 40% of enterprise applications to embed task-specific AI agents by the end of 2026, up from under 5% in 2025 — and, on the decision side, half of business decisions to be augmented or automated by AI agents by 2027.
Here is the part nobody is staffing for. The bottleneck will not be model quality. It will be identity, permissions, approval workflows, and audit. The plumbing. An agent that can issue a $42,000 purchase order is not a chatbot with ambition. It is a user with a credit card, and most companies have not decided who approves its spending.
I spend my days in enterprise CRM and ERP, and I can tell you the demo never covers this part. The demo shows the agent answering. Production asks who the agent is, what it may touch, and who says yes before the expensive actions run. In 2027, those questions stop being architecture trivia and become the job.
None of this plumbing is glamorous. An approval workflow is a policy that says this agent may spend up to here on its own, must ask a human between here and there, and may never do that third thing at all — plus a log that proves what happened when anyone asks later. Unglamorous. It is also the difference between an agent that is a tool and an agent that is a liability with a login.
3. The governance gap gets priced
Deloitte's numbers describe the gap precisely. In its 2026 State of AI in the Enterprise survey, 74% of companies expect to be using AI agents at least moderately by 2027 — but only 21% report a mature governance model for them, and just 11% have agents in production today. Adoption is sprinting. The controls are walking.
Gartner's prescription is proportional governance: classify agents by autonomy — observe, advise, act with approval, act autonomously — and match the controls to the level, instead of treating every agent as either locked down or fully trusted. Get it wrong in one direction and you strangle useful agents until teams build shadow workarounds. Get it wrong in the other and the first incident does your governance design for you.
That gap is about to get expensive, and the market has noticed. The companies that sold you the agents will happily sell you the guardrails you skipped the first time. The middle layer — the controls that sit between an agent and the systems it can change — is becoming a product category, because it has to.
Here is what that looks like in practice. In October, Transcend launched Rails, a policy-as-code platform that decides in real time, action by action, what an AI agent is allowed to do: plain-language rules, an accountable owner, a complete audit trail, a kill switch. It launched with live customers in media, consumer, and healthcare. Whatever you think of any single vendor, the shape of the product is the story. Governance is no longer a document. It is a runtime decision, made before the action executes.
Then the biggest enterprise software company on earth said the thesis out loud. On October 10, Microsoft CEO Satya Nadella published an essay, "Models as Insider Risks in the Super Intelligence Era," arguing that AI models should be handled the way security teams handle powerful insiders — not because they are malicious, but because any capable actor with access to important systems can make mistakes or be compromised. His prescription: "separate the supply of intelligence from the authority over it," with identity, privilege limits, activity logging, and containment all sitting outside the model, and an authorized person always able to pause or shut a model down mid-task. His hardest line is the one every deployer should sit with: "We simply can't outsource responsibility for what intelligence does on our behalf." When Microsoft's CEO tells enterprise buyers the controls must live outside the model, the middle layer stops being a startup's pitch and becomes a procurement requirement.
And the regulators are writing it down. On October 15, the Financial Stability Board — the standard setter for the world's banks — published final AI guidance telling firms to set "clear, risk-based triggers for when an agent must seek human approval." Read that carefully. The approval is not something the agent decides to seek. It is a trigger, defined in advance, by someone who is not the agent. The FSB also dropped its earlier suggestion that firms treat AI as "synthetic employees," replacing it with the acknowledgment that "AI agents do not reason like humans and may interpret pre-defined goals in unexpected ways." The HR metaphor is dead. The regulator now says what operators already know: the agent does not understand your intent. It executes its goals, and somebody human signs off before the risky ones run.
The US regulators are not waiting for the standard to trickle down. Reuters reported this month that the OCC, the Federal Reserve, and the FDIC are already asking banks what limits sit on their agent systems, how agent actions get monitored and approved, whether decisions can be audited, and how an autonomous system gets shut down. When the examiner's questionnaire asks how you approve an agent's actions, "the agent is built right" is not an answer.
So prepare for this: in 2027, governance stops being the slide at the end of the AI strategy deck. It becomes the budget line, the audit question, and — for the deployments that skip it — the cancellation reason. The companies still running agents when the cull ends will not be the ones with the best models. They will be the ones that built the controls early.
Next in this series: Part 2, "What to Learn Now" — the skills that compound in 2027, from agent evaluation and red-teaming to designing approvals a tired human will actually read.