The Companies Quietly Replacing Entire Workflows with AI Agents — While You’re Still Debating Prompts

There is a growing divide in the industry. On one side: teams shipping autonomous agent workflows that process claims, onboard clients, and run due diligence — in production, at scale. On the other: teams still running prompt engineering workshops. This piece is for the second group.

You have probably seen the headlines. Goldman Sachs deploying autonomous agents for trade accounting. JPMorgan running 450+ AI use cases. Lemonade settling insurance claims in three seconds. But here is what the headlines do not capture: the shift happening underneath.

The companies pulling ahead are not doing anything exotic with models. They are not using a secret version of GPT or Claude that you do not have access to. The models are the same. What is different is that they stopped thinking about AI as a tool you prompt — and started thinking about it as a worker you deploy.

That distinction changes everything.

The prompt trap

Most organizations are stuck in what we call the prompt trap. It looks like progress, but it is not.

The prompt trap works like this: a team discovers that an LLM can summarize documents, draft emails, or answer questions about internal data. They get excited. They build a proof of concept. Someone gives a demo to leadership. Leadership says “this is great, let’s do more.” So the team spends the next six months optimizing prompts, building RAG pipelines, fine-tuning retrieval, and arguing about whether to use GPT-4 or Claude.

Meanwhile, nothing changes about how work actually moves through the organization.

The claims adjuster still waits two days for the risk analyst. The financial advisor still coordinates four specialists over a week. The legal team still reviews contracts sequentially, one at a time, losing context at every handoff. The AI initiative produces a chatbot that answers questions — sometimes correctly — and a handful of people use it when they remember it exists.

This is not transformation. This is a science project with a Slack channel.

The companies pulling ahead recognized something that most teams have not: the bottleneck was never the model. It was the workflow. The ones that figured this out early are already redesigning entire operating models around agents — while everyone else is still optimizing prompts.

What the quiet movers actually did

The organizations getting real results did not start with “how can we use AI?” They started with “where does our process break?”

And the answer, in almost every case, was the same: at the handoffs.

They found the handoff that was killing them

Every complex business process has a moment where work passes from one specialist to another — and that is where time, context, and quality go to die. The risk analyst finishes their assessment and writes a summary. The market strategist reads the summary, misses a nuance, and builds their analysis on an incomplete picture. The tax specialist gets a different version of the story. By the time the portfolio manager tries to synthesize everything, they are reconciling three slightly different interpretations of the same client’s situation.

The quiet movers found their version of this handoff and asked a different question: what if the specialists worked in parallel, shared context automatically, and the handoff was not a handoff at all — but a structured collaboration? This is exactly what a multi-agent wealth management workflow looks like in practice — four specialist agents replacing four sequential human handoffs.

They deployed agents, not chatbots

The difference is fundamental.

A chatbot waits for a human to ask a question, generates a response, and stops. It is reactive. The human drives every step.

An agent takes a goal, decomposes it into steps, executes those steps using tools and data sources, coordinates with other agents, adapts when something unexpected happens, and delivers a result — with the human providing oversight at defined checkpoints, not micromanaging every action.

This is what Goldman Sachs built with Anthropic. Not a chatbot that answers questions about trade accounting — an autonomous agent that does trade accounting. Not a copilot that helps with client onboarding — an agent that runs client onboarding. The human reviews and approves. The agent does the work.

They invested in reliability before scale

Here is the part most teams skip — and it is why most pilots never reach production.

A demo can afford to fail. If your proof of concept crashes, you restart it. If the output is wrong, you adjust the prompt and try again. Nobody depends on it. The stakes are zero.

A production workflow cannot afford to fail. If your claims processing agent crashes at step 7 of 10, you cannot tell the policyholder “sorry, we need to start over.” If your underwriting agent produces an incorrect risk assessment, you cannot discover it three months later when a claim comes in. If your compliance workflow skips a step, you cannot explain to the regulator that “the AI had a bad day.”

The quiet movers understood this. Before they scaled a single workflow, they answered five questions that separate successful deployments from public failures:

1. If it crashes mid-workflow, does completed work survive? Not “we’ll add error handling later.” Right now, today — if the system goes down, do we lose everything or resume from the last checkpoint?

2. Is every decision traceable? Can an auditor — internal or external — reconstruct exactly why the agent made a specific recommendation, what data it used, and what alternatives it considered?

3. Can a human intervene at any point? Not just at the end. At any step where the stakes warrant it. Can the system pause, present its reasoning, and wait for human approval before proceeding?

4. Does it work when the model provider has an outage? If you are locked to a single model and that model goes down, does your entire operation stop? The companies that thought about this early route across multiple models — not for capability, but for resilience.

5. Can your agents talk to other systems? The industry is converging on standard protocols — MCP for tool use, A2A for agent-to-agent communication. Agents that cannot interoperate are agents that cannot grow beyond your first use case.

The teams that skipped these questions built impressive demos. The teams that answered them built production systems. (For the full checklist of what enterprise security looks like in agent infrastructure, we published a detailed breakdown.)

The gap is not what you think

Here is the uncomfortable math.

The technology gap between early movers and everyone else is essentially zero. Anyone can sign up for the same APIs, use the same models, and read the same documentation. There is no secret sauce in the model layer.

The gap is operational.

Morgan Stanley has been running AI-assisted advisory workflows since September 2023. That is over two and a half years of learning — which edge cases matter, where human oversight adds the most value, how to measure quality, how to earn trust from advisors who were skeptical at first. Every quarter their system runs, the organization gets better at deploying the next system. Not because the model improved. Because the people and processes around it improved.

JPMorgan started their AI journey with a specific loan processing workflow in 2017. Nine years of institutional learning. When they deploy agentic AI now, they are not starting from scratch — they are building on a decade of understanding how AI fits into regulated workflows.

This is the gap that widens.

Technology gaps close — models improve, prices drop, tools get easier. But operational gaps compound. Every quarter an organization runs agent workflows in production, they accumulate knowledge that cannot be purchased, licensed, or copied. It lives in the judgment calls of the people who operate the system. It lives in the workflow designs refined through thousands of real-world runs. It lives in the institutional confidence that lets a team move from “let’s pilot this” to “let’s scale this” without a six-month committee review.

The companies debating prompts today are not just behind on technology adoption. They are falling behind on operational maturity — and that is the gap that is hardest to close.

The three-month window

We are at a specific moment in the adoption curve where the dynamics are about to shift.

Right now, deploying AI agents in production is still a differentiator. The majority of organizations are still in evaluation, still running pilots, still waiting for “the right moment.” If you deploy a working agent workflow today, you are ahead of most of your industry.

In twelve months, that will not be true. The tooling is getting easier. The patterns are becoming well-documented. The enterprise playbooks are being written in real time by early movers like Goldman, JPMorgan, and Salesforce. The barrier to entry is dropping fast.

What will not be easy to replicate in twelve months is the operational maturity that comes from running these systems now. The edge cases you discovered. The workflow refinements you made. The trust you built with your team. The institutional knowledge that turns a working prototype into a competitive advantage.

This is the three-month window. Not because the technology will change — but because the cost of waiting is compounding.

Every quarter of delay is not just lost time. It is lost learning. And learning compounds.

What “starting” actually looks like

If you have read this far and recognize your organization in the “debating prompts” description, here is what moving forward actually looks like. Not a twelve-month roadmap. Not a new department. Three concrete actions in the next thirty days.

Week 1: Find your most expensive handoff

Walk through your highest-volume workflow. Not the most complex one — the one that runs the most often and touches the most people. Map it on a whiteboard. Mark every point where work passes from one person to another. At each handoff, estimate: how long does the next person wait before picking it up? How much context is lost in the transition? How often does the handoff introduce an error or inconsistency?

The handoff with the highest combination of wait time, context loss, and error rate is where you start. Not because it is the most strategic — because it is the most wasteful. And waste is easy to measure, easy to justify, and easy to demonstrate improvement against.

Week 2: Build one agent that replaces one handoff

Not the whole workflow. One handoff. If your claims process has five specialist steps, pick the one where the most time is lost. Build an agent that does what that specialist does — pulls the right data, performs the analysis, produces the output that the next step needs.

Keep the human in the loop. The agent does the work. A human reviews and approves. You are not removing anyone — you are removing the wait, the context loss, and the inconsistency.

Week 3-4: Run it alongside the existing process

Do not replace anything yet. Run the agent workflow in parallel with the human process. Compare outputs. Measure accuracy the way you would measure any software system. Track time savings. Let the team see what the agent produces and flag where it gets things wrong. This is not testing — this is teaching your organization what working with agents feels like. The confidence that comes from seeing an agent produce the right answer, week after week, is what turns skeptics into advocates.

The question is not “should we?” anymore

The evidence is in. Goldman Sachs, JPMorgan, Morgan Stanley, Salesforce, Lemonade, Harvey AI — these are not speculative bets on unproven technology. They are production deployments generating measurable returns, run by organizations with the most to lose from getting it wrong.

The question has shifted. It is no longer “should we deploy AI agents?” It is “how quickly can we build the operational muscle to deploy them well?”

And that question has a time component. The companies that start now will have a year of operational maturity by the time the rest of the industry catches up on tooling. They will have discovered the edge cases, refined the workflows, built the team confidence, and earned the institutional trust that turns technology into competitive advantage.

The companies that are still debating prompts will eventually ship their first agent workflow. By then, their competitors will be on their tenth.

Start building

The infrastructure patterns the leaders share — crash recovery, audit trails, human-in-the-loop, model flexibility, agent interoperability — are the same whether you are Goldman Sachs or a ten-person team. The difference is that Goldman spent six months and embedded engineers building it from scratch.

You do not have to. This is why we built JamJet.

JamJet is an open-source runtime (Apache 2.0) that provides the production foundation these workflows require: event-sourced durability, automatic execution traces, first-class human approval gates, multi-model routing, and native MCP + A2A protocol support. Rust core, Python SDK. The infrastructure patterns the industry leaders spent millions building — available with pip install jamjet.

• See what your competitors are already doing →
• Read the full business case for AI agents →
• Watch a multi-agent workflow in action →
• Try JamJet — pip install jamjet →
• Read the docs →
• Join the community →