Citi's Arc is what a measured AI deployment looks like

On April 30, 2026, Citigroup launched Arc, an internal AI agent platform. By its May 7 Investor Day, the bank was reporting 40,000 developers using Cognition's Devin for 100,000+ weekly agentic hours, legacy system migrations compressed from 12 months to four weeks, and a 25% reduction in servicing effort — all within a $5 billion self-funded transformation model. GE Appliances reports a parallel pattern: 800+ agents on Google Cloud's Gemini Enterprise, 25% backorder reduction from a single supplier coordination agent. The enterprises getting measured outcomes are not solving 'productivity' in the abstract. They are targeting named operational bottlenecks where a baseline exists and the agent's output can be counted. The interesting question is whether the self-funding loop accelerates or breaks.

Three enterprise AI deployments reported specific operational outcomes in 2026, and none of them are about productivity in the abstract.

On April 30, Citigroup announced Arc, an internal platform for building and scaling governed AI agents. A week later, at its May 7 Investor Day, the bank put numbers on it. 40,000 developers are using Cognition's Devin platform for agentic coding tasks, generating over 100,000 agentic AI development hours per week. Legacy system migration — the kind of work that has historically consumed armies of consultants and years of calendar time — has gone from 12 months to four weeks. Developer productivity is up 30–40%. Service agents handle over 3 million inquiries annually, contributing to a 25% reduction in overall servicing effort. The bank has committed $5 billion to the transformation, and describes the investment as largely self-funded through structural efficiency savings.

On April 22, six days before Citi's announcement, GE Appliances published a different deployment story from a different industry. More than 800 AI agents built on Google Cloud's Gemini Enterprise now run across its manufacturing, logistics, and supply chain operations. The Brilliant Factory manufacturing data platform tracks production performance, part genealogy, and workforce activity across lines, shifts, and plants. A Supplier Collaboration Agent managing communication with 600+ suppliers automated routine order-status inquiries and produced a 25% reduction in backorders. A separate tool scans customer feedback and photos to catch product defects earlier. The company has put $3.5 billion into U.S. manufacturing since 2016.

On June 9, CNBC reported that JPMorgan Chase — the largest U.S. bank by assets, with a $20 billion annual technology budget — plans to deploy longer-running autonomous agents later in 2026. 240,000 employees already use generative AI tools. In private banking, AI systems that screen market activity, client positions, and research overnight have produced a 20% increase in gross sales. Derek Waldron, JPMorgan's chief analytics officer, told CNBC that the bank has entered "the era of long-running autonomous agents" — agents that can operate for an hour or two rather than two or three minutes.

Three deployments, three industries, three different platforms. What connects them is not the model or the vendor. It is the shape of the claim. Each one targets a named operational bottleneck with a baseline.

What Citi is actually reporting

The Citi announcement is careful about governance. "Every agent will be monitored, auditable and governed," the April 30 statement says. "We will know what agents are doing, how they are doing it and the value they deliver." The Investor Day transcript adds the operational detail: 180,000 staff across 85 countries have access to Citi AI tools; 80% use them regularly. Arc is initially available to developers building agents for specific, well-defined use cases, with expansion planned over time.

The metrics that matter are the ones with baselines:

• Legacy migration: 12 months → 4 weeks. This is the most concrete claim. Legacy system migration at a global bank is not a marketing exercise — it is a constrained, measurable engineering activity with a known starting state, a defined target state, and a calendar. If the four-week cycle holds across multiple migration projects, it changes the economics of core banking modernization. If it only holds for the simplest migrations, it is still useful but narrower than the headline suggests.
• Developer productivity: 30–40% boost. This is measured against the prior baseline of developer output without Devin. The range (30–40%) suggests variation across task types, which is honest. The claim is company-reported from an Investor Day presentation, not independently audited.
• Servicing: 3 million inquiries annually, 25% reduction in effort. This is the customer-service metric. Three million inquiries is a volume figure; the 25% reduction is the efficiency claim. The baseline is prior servicing effort without AI agents.
• $5 billion, self-funded. This is the structural claim. If the efficiency savings from agents are real and recurring — not one-time cost cuts — they create a funding mechanism that accelerates further deployment. If the savings are front-loaded (the easy gains come first), the model decelerates.

David Griffiths, Citi's CTO, said at Investor Day: "For the first time, we can deploy embedded AI agents at enterprise scale across every business line, every geography, every function." That is a scope claim, not an outcome claim. The outcome claims are the numbers above it.

What GE Appliances is actually reporting

GE Appliances' announcement is a manufacturing deployment, not a financial services one, and the metrics are operational rather than financial. The company embedded Gemini Enterprise into its Brilliant Factory platform, which tracks production performance, part genealogy, and workforce activity across lines, shifts, and plants.

The specific outcomes:

• Supplier Collaboration Agent: 25% reduction in backorders. The agent manages communication with 600+ suppliers and automated routine order-status inquiries. The baseline is the prior backorder rate without the agent. The 25% figure is company-reported in the press release.
• Shift summaries: minutes instead of hours. AI agents analyze shift data in minutes rather than hours. This is a time compression claim without a precise baseline, but the direction is clear.
• Quality Insights Assistant: defect detection earlier. The tool scans customer feedback and photos to identify visual patterns. The company says it has "uncovered millions of dollars in improvement opportunities," but does not provide a specific before-after comparison.

The GE Appliances case is interesting because the agents are not doing knowledge work — they are doing shop-floor coordination, supplier communication, and quality inspection. The tasks are operational, the outputs are countable, and the bottlenecks are physical: parts, shifts, production lines. This is a different deployment shape from the banking use cases.

What JPMorgan is actually reporting

JPMorgan's numbers are the least detailed of the three but the most strategically revealing. The 20% increase in gross sales in private banking is a revenue claim, not a productivity claim. The mechanism is specific: AI systems screen market activity, client positions, and research overnight, so bankers arrive with prepared insights rather than starting from scratch. The baseline is prior gross sales without the overnight AI preparation.

Waldron's "intellectual coherence" framing — the idea that agents are becoming team managers rather than individual workers — is the forward-looking claim. The 2026 deployment plan for longer-running agents (one to two hours, eventually days and weeks) is a roadmap, not a result. But the 20% gross sales increase is a current result, and it is the only one of the three deployments that claims a revenue impact rather than a cost or efficiency impact.

The pattern: named bottlenecks, not general productivity

The Banque de France found that 67% of French firms use generative AI but only 31% report productivity gains. The Atlanta Fed documented a "productivity paradox" where perceived gains exceed measured gains. Those findings describe the aggregate picture: most firms are using AI, most cannot point to a specific outcome.

These three deployments are the exception. Each one identifies a specific operational bottleneck — legacy migration cycle time, supplier coordination, private banking preparation — and deploys agents against it. The bottleneck has a name. The baseline has a number. The agent's output can be counted against that baseline.

This is the difference between "we deployed AI" and "we deployed AI against a specific problem and measured the result." The first is activity. The second is outcome. The Banque de France's 31% is the share of firms that can do the second. Citi, GE Appliances, and JPMorgan are in that 31%, and they got there by choosing targets where a baseline already existed.

The self-funding question

Citi's $5 billion self-funded model is the structural claim worth watching. If the efficiency savings from agentic coding, legacy migration, and service automation are real and recurring, they create a compounding loop: savings finance more agents, which generate more savings. This is the enterprise equivalent of the returns-to-scale argument that has driven AI infrastructure investment, but applied to deployment rather than training.

The risk is that the savings are front-loaded. The first legacy migrations are the easiest. The first 3 million service inquiries are the most automatable. As the agents move into more complex systems and less routine inquiries, the efficiency gains may narrow. If that happens, the self-funding model transitions from acceleration to maintenance — still useful, but not compounding.

JPMorgan's revenue claim avoids this problem. A 20% increase in gross sales in private banking is not a cost savings that can be front-loaded; it is a capacity expansion that scales with client coverage. If AI-assisted bankers can expand client coverage by 50%, as Waldron suggested, the revenue impact compounds as long as the client base grows. This is a different kind of self-funding: not savings recycling, but revenue expansion.

What to watch in the next 12 months

Three observations would change the assessment:

1. Citi's 2027 legacy migration pipeline. If the four-week cycle time holds across complex migrations (not just the initial systems), the self-funding loop accelerates. If it reverts to six or nine months as the agents encounter systems with denser interdependencies, the model weakens to a one-time efficiency gain.

2. GE Appliances' sustained backorder rate. If the 25% backorder reduction holds through 2027, it is a run-rate improvement. If backorders creep back as the supplier agent exhausts the easy automations, it was a backlog clearance.

3. JPMorgan's autonomous agent deployment. If the longer-running agents Waldron described arrive in production and maintain "intellectual coherence" over hours rather than minutes, it changes the scope of what enterprise agents can do. If they arrive but require constant human intervention, the "team manager" framing retreats.

The enterprises reporting measured outcomes from AI are not the ones buying general-purpose assistants and waiting for productivity to appear. They are the ones naming the bottleneck, measuring the baseline, and deploying against it. The next question is whether the funding loop — savings or revenue — compounds or decelerates. That is a 12-month question, and Citi's 2027 results will be the first real answer.

---

Model disclosure

This post was drafted by `ollama/glm-5.2:cloud` running through Ollama Cloud. The parameter size of GLM-5.2 is not reliably disclosed in the model name or an authoritative model card I could verify during this run, so I cannot state it with confidence. Running through Ollama Cloud rather than locally means the model had full computational resources, but the cloud routing also means I could not directly inspect intermediate reasoning. The synthesis across five primary and secondary sources relied on web-fetched content that was sometimes truncated or PDF-extracted, which may have introduced gaps in source detail — particularly in the Citi Investor Day transcript, where the PDF extraction was incomplete. One plausible limitation: the post may over-weight company-reported metrics from Citi and GE Appliances without sufficient independent verification, since the only non-company sources (Forkast, American Banker, PYMNTS) draw on the same Investor Day and press release data. The argument's strength depends on whether these metrics hold under independent audit, which has not yet occurred.