Labs / Research · Updated 2026-10-14

Your company's most valuable knowledge may not be in its software

Most companies can list their software. Few have written down how the work actually gets done. AI is about to make that gap expensive.

By Founder

Published

PlannedDecided, not built.

Explore with AI

Field report

01

The procedures people actually follow rarely match the manuals

Ask a business what software it uses, and you will get a list: accounting, customer management, purchasing, document storage, inventory and a collection of specialized industry tools. Ask how an order becomes an approved quote, and the explanation gets complicated. Someone knows which fields can be trusted, which manufacturer rules supersede a brochure, and when a supervisor needs to step in.

That knowledge is often carried by experienced employees rather than by the software itself. The system stores a completed transaction. It may say little about the phone call, the exception, the rejected option or the judgment that preceded it.

A transaction record shows what happened. Operational knowledge explains why.

02

Observation is becoming a software input

Microsoft Power Automate task mining can capture desktop actions and reconstruct routine tasks so a company can understand where work is duplicated or interrupted. Separately, Copilot Studio's computer-use capability allows an agent to interact with applications through their visible interface. That matters when a business relies on desktop software or vendor systems that lack suitable APIs.

Those capabilities solve different problems. One observes a person's workflow. The other executes instructions in a computer interface. Joining them with persistent memory and models that can reason over demonstrations suggests a new way to document and assist work. It does not mean an agent automatically learns an entire occupation from a screen recording.

Capturing steps and reliably performing decisions are different technical milestones.

03

What a screen recording cannot tell you

Consider a configured commercial window order. A purchasing specialist compares dimensions, glazing, finish, hardware and manufacturer documentation before preparing a quote. A screen recording can show every field entered. It may not reveal that the specialist recognized an obsolete document, understood a sizing exception, or called a manufacturer representative to confirm a condition that the software could not validate.

A working agent needs the rules behind the click path: the authoritative source for product options, the approval hierarchy, what information is too old to rely on, and when the answer must be ‘unknown’ pending verification. If those rules are missing, repeating the demonstrated action can reproduce the error more efficiently.

Imitating an action is easier than recovering the judgment that made it correct.

04

The computer is being reorganized around agents

The familiar computer put an employee in front of applications. An emerging agent-driven workflow places an orchestration layer between the employee's objective and the applications that execute it. Models interpret requests; the harness coordinates memory, tools and safeguards; connected applications remain the systems of record.

The popular analogy of model as CPU, harness as operating system and skills as apps is useful shorthand. It is not a literal systems diagram. The missing layer is business context: a company's current products, permissions, process rules, locations, authorized claims and exceptions. Without that information, a sophisticated agent can perform a technically successful sequence that produces the wrong commercial result.

Business context is part of the operating architecture, not a paragraph pasted into a prompt.

05

A model-independent workflow is a business decision

Different tasks justify different models. Complex document interpretation may demand one model; routine extraction and categorization may be better handled by a smaller one. The orchestration system should preserve the workflow's rules and evidence so an organization can change models without rebuilding its operating knowledge.

That flexibility is valuable, but model routing by itself is easy to imitate. The harder asset is a maintained map of how a particular industry's work gets done, where facts come from, what an agent is allowed to do and how exceptions get reviewed.

Keep the workflow's truth and controls outside the model.

06

The right sequence for a real business

Start with one process that employees can describe from beginning to end. Capture it with consent. Identify which inputs are stable and which change daily; who owns each field; what decisions staff make outside the software; and which approvals are mandatory. Then document failure cases before automating a single external action.

An agent may first read a document, organize a draft or prepare a comparison for review. It earns broader authority only after its work can be checked against the same source and acceptance criteria the business already trusts. An application with no API can still be operated through a screen, but desktop access does not confer permission to place orders, expose customer records or authorize spending.

Capture, verify, assist, review, then expand the allowed actions.

07

From unwritten procedure to Business Truth

At Lehvel, we use Business Truth to describe the checked facts a business relies on and publishes: what it is, what it offers, where it operates, what makes its claims credible and which source is authoritative. Internal workflow knowledge belongs beside that record. It adds operational truth: who can approve a quote, which system controls the latest product configuration, and what an automated process must stop to verify.

As agents begin handling more work, the distinction between a web presence and an operating process becomes less useful. The information that powers an accurate answer to a customer should connect to the information that lets an employee or agent act on that answer correctly. Publishing a product page, choosing its source data and controlling the transaction are different parts of one system.

The objective is a business the software can understand accurately enough to assist, and whose people remain accountable for consequential actions.

08

What remains unresolved

Workplace observation raises real questions about consent, confidentiality and employee trust. A useful workflow study should not become continuous employee surveillance by default. Screen captures may contain personal information, customer records or credentials, and storage and access need clear limits.

Reliability is also unsettled. Stanford's 2026 AI Index documents striking progress on structured computer-use benchmarks while recording substantial remaining failures. Those results are a warning against treating a short demonstration as evidence of safe end-to-end autonomy. The next commercial advantage may belong to companies that document decisions and exceptions before connecting another model.

What we asked and what we found

Every entry answers the same eight questions, in this order.

Question

How much of how a company works can AI learn by watching, and what has to be written down before AI can do the work reliably?

Why it matters

Tools that record desktop work and tools that operate a computer for you are arriving together. The business question is whether they capture the decisions that make work correct, or only the clicks.

State

Planned Decided, not built.

What we built or tested

  • A desk review of what task-recording tools, computer-operating AI and published test results can and cannot do, drawn from vendor documentation and an academic survey.
  • A plain sequence for a business with older software: watch with consent, write down the steps, ask an experienced person for the rules, confirm the facts, automate one narrow task, review the exceptions.
  • We did not record any employee, and we claim no productivity result.

Evidence

Task mining capability (Microsoft Power Automate; read October 14, 2026)
Records and analyzes desktop task steps to show how people execute workflows and where automation might help.
Computer use for enterprise applications (Microsoft Copilot Studio; read October 14, 2026)
Agents can use graphical user interfaces, including desktop applications with no direct API, under configured access and execution controls.
Computer-use benchmarks (Stanford HAI 2026 AI Index; read October 14, 2026)
Structured benchmark performance has improved rapidly but retains material task failure rates. Benchmark success is not evidence that an agent can replace an employee.
Desktop access model (Anthropic Cowork documentation; read October 14, 2026)
Computer access requires permissions and can expose the agent to whatever sensitive material is visible on-screen.

What we learned

  • A recording shows which screens someone opened. It cannot show why an unusual order was approved or a substitute rejected.
  • In older industries, the handoff between a person's judgment, the catalog, the quoting rules and the accounting system matters more than any single program.
  • The model can change. The checked facts, the rules and the approvals have to live outside it.
  • Start with a narrow job that has clear inputs, a visible result, an exception path and a person with authority to settle doubt.

Limitations

  • This is desk research and our own interpretation. We did not study any customer's employees.
  • We found no basis for the claim that watching someone for a week lets AI do their whole job within a year.
  • A vendor demo or benchmark does not show how AI performs on real work, with changing screens, exceptions, privacy rules and restricted logins.

What changes next

Study one repeatable job with a willing operator, with consent. Write down its decisions, sources, exceptions and permitted actions, then compare AI-assisted work with the review already in place.

Related

The building-materials product-information chainBusiness Truth researchRead more research and field notes
Failed tests are published as results. A state on this page describes today, not plans.

Sources and credits

These tools belong to the people who made them. Lehvel reviewed and, where noted, tested them. Nothing here is a Lehvel product, and no performance figure attributed to a vendor has been reproduced by us.

Next in Labs

← All of Labs

Explore with AI

Get Upstream

One email a week: what we asked the AI, what it got wrong, and how we checked.

Our research reaches 17,000+ followers across our channels. Upstream is the weekly version.

Weekly. Unsubscribe anytime. Privacy