RPA Automation Anywhere is a recorder plus a runtime. The recorder format is where the architecture cracks.

2. What Object Cloning actually serializes.

Below is a representative serialized element fingerprint. The property names match what the vendor documentation describes the recorder as capturing for the STANDARD technology (Windows controls). The shape is the point: pixel-adjacent coordinates, a path token, a state token, and the visible name all travel together as a single match unit.

<!-- Automation Anywhere Object Cloning: serialized element fingerprint -->
<!-- Captured by Bot Creator. Replayed at runtime by Bot Runner. -->
<element>
  <technology>STANDARD</technology>
  <name>Post invoice</name>
  <path>Toolbar/Group/Button[3]</path>
  <left>847</left>
  <top>234</top>
  <width>92</width>
  <height>28</height>
  <state>Focused, Enabled</state>
  <value></value>
  <html_innertext></html_innertext>
</element>
<!--
A vendor UI release that nudges the toolbar three rows down,
inserts a validation step in front of the button, or renames a
path segment will invalidate this fingerprint at runtime.
The bot raises 'Unable to perform Click action (TYPE_024)'
and a developer has to re-record the element.
-->

A click on a static button on a static page works fine in either format. The difference shows up the first time the underlying surface changes. The fingerprint on the left has roughly ten independent ways to invalidate; the match on the right has two (role and name), and a screen-reader user would notice exactly the same two changes if either one drifted.

3. The other recorder format, in source.

The recorder Mediar ships does the opposite of what Object Cloning does. It treats the pixel stream as noise and only serializes events whose meaning survives a layout change. The function below is the literal filter, copied from the desktop agent’s source file. Six event types pass through. Everything else is dropped before the workflow file is authored.

Three things are worth pointing at in that filter. First, there is no mouse_move, scroll, or raw_keystroke case; those are explicitly dropped at the source rather than recorded and post-filtered, because keeping them around is what tempts a recorder to identify elements by where the cursor went. Second, text_input_completed is the aggregated final value after a focus loss, not a stream of keydown events; the recorded artifact says “the user typed 30 March 2026 into the date field,” not “the user pressed 3 then 0 then space then M….” Third, the file path is stable: you can grep for is_meaningful_event_type in the public Terminator SDK on github.com/mediar-ai/terminator and read the same six-event filter.

4. The runtime: how a recorded artifact reaches the OS.

Automation Anywhere’s runtime is Control Room (orchestration, credentials, scheduling, audit) plus Bot Runner (the local agent on the machine that hosts the target app). A bot is dispatched from Control Room to a Bot Runner, the runner loads the recorded task, and at every step it asks the OS “find me the element that matches this fingerprint.” The accessibility-API runtime has a similar shape, but the question it asks the OS is different: “find me the element with this role and this name.” The diagram below traces the second shape end to end.

The model in this pipeline lives at authoring time (when the workflow file is generated from the recording), not at runtime. That matters for two reasons relevant to anyone replacing Automation Anywhere. The runtime stays deterministic, so the compliance team can read an audit log without needing to know what a model decided. And the cost of a step is the cost of an OS call (microseconds), not the cost of an LLM call (hundreds of milliseconds plus tokens).

5. The economic shape that follows from the format.

A platform that captures roughly twenty properties per element spends most of its operating cost in two places: license fees that price the runtime per bot, and people who keep re-pointing the recorder when fingerprints invalidate. A platform that captures two properties per element does not have the second cost in the same shape, because most UI shifts do not invalidate a Role+Name match. The numbers below are typical of what an enterprise RPA buyer is comparing in 2026.

The 70% number is from a specific deployment (an LG-customer F&B chain that moved a SAP B1 order-entry workload off UiPath onto Mediar; the same architectural argument applies to an Automation Anywhere migration on the same surface). The other numbers are the shape of the published vendor pricing and the literal property counts in the two recorder formats. None of them are a press-release quote and none are computed against an invented baseline.

6. The counterargument: when Automation Anywhere still wins.

Three cases. The first is when the workflow’s binding constraint is integration breadth into the Automation Anywhere ecosystem itself: the IQ Bot pre-built document classifiers, the Bot Store catalog, an existing center of excellence with sunk training cost, or a procurement office that needs the buyer’s existing master agreement. Those are real pieces of value and they do not transfer.

The second is when the target application exposes nothing useful through the OS accessibility tree. Some custom-rendered Direct2D canvases, certain remote-rendering stacks where the UI tree is empty on the local machine, and a small number of specialty industrial control surfaces fall into this bucket. On those, the best you can do is fall back to a vision-based recorder, which is a different category of bet from either Object Cloning or an accessibility-tree match.

The third is when the buyer has already spent enough on Automation Anywhere implementation that the sunk cost outweighs the maintenance cost the format imposes going forward. That is a financial argument, not a technical one, and it has an expiration date that the maintenance backlog itself eventually moves past.

7. Resolution: pick the format that survives a UI release.

The decision is at the recorder layer, not the dashboard layer. Two questions, in order. Does the workflow live on a surface that exposes a usable accessibility tree (SAP GUI, Oracle EBS, Jack Henry, Fiserv, FIS, Epic, Cerner, Office, most Win32 desktop apps, and any modern web app do)? If yes, an accessibility-API recorder will absorb most of the UI shifts your Object Cloning bots currently break on. Does the workflow run often enough that the maintenance backlog is a real cost line (more than ten times a week, or with a SLA on completion)? If yes, the format choice is load-bearing for the next 24 months of operating cost.

If the answer to either is no, the architectural argument here is weaker and a feature-list comparison is the right next step. If the answer to both is yes, the migration shape in the FAQ below (one workflow, two-week parallel run, cut over on zero divergence) is the cheapest way to find out whether the recorder format actually does what this page says it does on your specific surface.

Common questions about RPA Automation Anywhere