Work Done vs Workflows: Delivering Results vs Building Workflows

Buying automation tooling and getting operational work done are two different purchases. Most teams pay for the first and still wait for the second.

The gap is measurable. McKinsey's latest State of AI survey (roughly 2,000 organizations across 105 countries, published November 2025) found that 88% of organizations now use AI in at least one business function. Only 39% report any earnings impact at the enterprise level, and most of those put it below 5%. Adoption is not the problem; adoption went fine. The work the tooling was supposed to absorb is still sitting in the queue.

The standard diagnosis is an execution problem: the rollout needs more time, the team needs more training, the next model will close the gap. We think that diagnosis is wrong. Every generation of automation software has shipped the same thing (a better way to build workflows) and left the same thing behind (responsibility for the work itself). This post is about that gap: why it survives every platform shift, why the only escape has been renting effort by the head, and what it looks like when the thing you buy is the work.

For a decade, software has delivered process

Look at what each generation actually shipped.

Classic RPA (robotic process automation) sold rule-based bots. What buyers got was a studio license and a construction project: encoding every process into scripts, then keeping those scripts alive as portals changed and formats drifted. A Forrester Consulting study on RPA scalability found that 45% of firms experience bot breakage weekly or more often, mostly for lack of support and maintenance. The license was never the expensive part. The bot was never really the product; the upkeep was.

Copilots and agent platforms changed the technology and kept the shape. The platform provides capability (models, orchestration, connectors) and you still translate your business process on top: the prompts, the integrations, the guardrails, the exception handling, the checking of whether outputs are actually right. McKinsey's own data shows where that leads. The roughly 6% of organizations it classes as AI high performers share one distinguishing behavior: they redesign workflows around AI rather than dropping tools into existing processes. Read that from the buyer's side. The winners are the ones who did the most work themselves. The platform didn't close the gap between tooling and results; additional labor did.

Vertical agents move the line, not the ownership

The newest generation looks like it solves this. A vertical agent (a customer support agent, a voice sales agent) arrives with a baseline process built in. That is a real improvement over a blank platform.

But look at where the responsibility sits. Training the agent on your policies, tuning it on your data, and above all owning the quality of what it produces: still yours. The vendor's SLA (service-level agreement, the contractually guaranteed level of performance) covers whether the agent is up. Whether the agent was right is your problem, and in an operations context, being right is the entire job.

The work under a real request rarely fits inside the conversation, either. In our banking deployments, answering one operational email touches five systems (the customer CRM, internal APIs, the data lake, the service desk, the inbox itself). A baseline agent that handles the dialogue still leaves the lookups, the validation, and the follow-through exactly where they always were.

The market is noticing. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, and its stated causes are escalating costs, unclear business value, and inadequate risk controls. Notice what is not on that list: model capability. The expected failures are ownership failures, questions of who pays, who measures value, and who controls risk. In the same research, Gartner estimates that of the thousands of vendors claiming agentic capability, only about 130 are real. If you have been skeptical of agent pitches, you have been reading the market correctly.

The escape hatch has been headcount

There has been one reliable way to buy guaranteed output quality: engage a vendor on a headcount-based service contract. It works, in the narrow sense that someone else contractually owns the result. It is also expensive, slow to change, and it scales linearly: double the volume, double the invoice. The process knowledge compounds in the vendor's staff rather than in anything you own.

Institutions did not choose this model because they liked it. They chose it because it was the only contract on offer where the deliverable was work done rather than a tool for doing work.

The SLA that matters

Nobody's SLA is measured in workflows built. It is measured in work done.

An ops leader's own commitments are not phrased in deployments or integrations. They are phrased in cases closed, responses sent, exceptions resolved, within a deadline and to a quality bar. Automation that is bought and sold in any other unit leaves a translation gap, and the buyer has always been the one holding it.

Patched closes that gap by taking the other side of the contract. We use agents to build, monitor, and maintain your workflows, and the SLA covers not just availability but output quality.

Your process (documents, SOPs, policies)
          ↓
        [Build]    encoded as a structured workflow, live in 2 to 5 weeks
          ↓
        [Run]      structured execution: validation steps, human review paths
          ↓
        [Monitor]  outputs measured against the quality SLA, continuously
          ↓
        [Maintain] drift found and fixed as part of the service, not as a ticket

The structure underneath is the same architecture we have argued for before: discrete steps, explicit checkpoints, uncertain cases routed to people with the reasoning logged. That architecture is what makes a quality SLA possible at all. You cannot guarantee what you do not measure, and you cannot measure a black box.

Ownership stays on your side of the line. The workflows are yours, and so is the choice of AI model; if we stop earning the contract, you keep both. Pricing follows the same logic: execution-based, no per-seat licensing, roughly 60% cheaper than the alternatives, because we are paid in the same unit you are measured in.

What our numbers look like

The pattern behind the numbers is the point. One email operations deployment now handles three times its original volume at 65% lower overall cost, and the analyst effort that used to go into lookups and evidence-gathering (roughly 75% of the manual work per case) now goes to the escalations and investigations that actually need judgment. Volume went up; cost went down; the people moved up the stack.

None of it trades away control. Every action carries a full audit trail (what was checked, what was sent, when, and by whom), and deployments run in a dedicated VPC (virtual private cloud) or on the institution's own premises, with role-based access, in line with RBI guidelines and SOC2 controls. For a regulated buyer, that is not a feature list; it is the difference between automation you can defend to a regulator and automation you have to explain away.

The honest summary

What buying work done does not mean is worth stating plainly. It does not mean autonomy; we have written before about why free-form agents fail in regulated operations, and nothing here changes that argument. Edge cases still reach people, by design, with the data gathered and the reasoning logged. Accountability does not transfer either: the institution owns every output, which is exactly why every step has to be auditable. And a quality SLA is only as honest as the measurement behind it (structured logs, baselines, regression tests); a vendor offering one without that infrastructure is guessing at your expense.

What does change is what the purchase is. For a decade, the market has sold better and better ways to build workflows, and the queue has never noticed. Tooling budgets went up; the backlog did not come down; McKinsey's 88-to-39 gap is that decade on a chart. The queue moves when someone owns the work.

Patched builds, monitors, and maintains structured AI workflows for regulated operations, with SLAs on work done, not just uptime. If your team is paying for automation and still waiting for the work, we're easy to find.