Class A+0.5 pts
Single agent
One agent in the middle. Input on the left, output on the right. It calls an LLM, makes a tool call, retrieves knowledge, and answers.
ExamplesMicrosoft Copilot agents, Lyzr Studio Agent
For enterprise teams starting their AI adoption One automation at a time
How enterprises adopt AI: the cumulative score.
Every agentic automation you put live adds to the same score, weighted by its class, and the next one starts from that higher base. Nobody reaches the end state in one program. You get there by adding one automation to the last, until the sum is the transformation.
One agentic automation, liveClass A, +0.5 ptsClass B, +1 ptClass C, +2 ptsThe cumulative AI adoption score
Illustration: 100 automations across one company. The score only goes up.
Before the flow, the premise
Understanding agent classes.
Not every agentic automation is the same size. Lyzr groups them into three classes, the same classes behind agent-month pricing: Class A retrieves, Class B acts, Class C is privileged and always on. Each class adds a different weight to the cumulative score.
Class B+1 pt
Multi-agent workflow
Agent 1 does something and passes it to agent 2, which passes it to an if-else. Two agents sit after the branch, and the workflow ends in one output.
Examplesn8n, Lyzr SuperFlow
Class C+2 pts
Always-On Agent
An agent that never stops. It runs its own coding agent inside, has skills and a memory (brain), and all of it is expressed as a file system it can read and write.
ExamplesGrokbot, Muse, Lyzr GitAgent
Section 1, one problem, one track
Take one problem at a time. Launch a track. Run it to auto mode.
A track is how one problem becomes one agentic automation. The same seven steps every time. Deploying is not the end: after deploy the track iterates on live feedback until the corrections fade, and then it runs in auto mode. The next track starts alongside it. That overlap is where the cumulative score comes from.
- 1Identify the problem
The owner names the pain and the number that should move.
- 2Discovery
Map the current process: steps, data, decisions, guardrails, who touches what.
- 3Build the automation
Agents, workflow and integrations, built in the room with the team.
- 4Test thoroughly
Evals, edge cases, guardrails, and sign-off from the people who own the process.
- 5Deploy
Into the workspace, next to the humans who run the rest.
- 6Iterate
Live feedback flows back into the build. Corrections become rules the memory keeps.
- 7Auto mode
The automation runs on its own. People stay informed, and iteration continues in the background.
One track at a time in place ofthe big-bang programthe 18-month roadmapthe pilot that never shipsslideware
Scroll sideways to follow the flow.
Where the track is right now
Done
Live feedback loop, iterate back to build
Live in production
Section 2, phases and the growing memory
Each phase is a collection of tracks. Human-run processes give way to agents, and the memory grows.
Imagine a team of 24 people running 24 processes. In phase one, a couple of them move to agents. Phase by phase, blocks shift from human-led to AI-assisted, AI-led and autonomous. By phase four, roughly 70% of the processes run on agents, and the memory (brain) on the right is far richer than when you started.
Human-led: a person starts it; agents answer when asked
AI-assisted: a person runs it; agents prepare the analysis alongside
AI-led: agents run it overnight and wait for approval or feedback
Autonomous: agents do the job and keep people informed
Phase 112% of the team's processes run with agents3 of 24 processes touched. The first couple of tracks, and the memory has started.
Illustration for a compensation team. The categories and the counts are set per phase to tell the story; the real split comes out of discovery.
Section 3, what this leads to
A thriving organizational memory (brain) that leads to emergent AGI behavior.
The memory (brain) now holds how your team decides, what it learned, and why. Given that context, agents start recommending the next best action for the team. Some actions they hand to a person. Some they execute themselves by launching subagents. Not a bigger model. A richer memory about your business.
Organizational memory (brain), phase 4
built by 17 tracks across four phases400 entries, 14 sources
Moved out of people's heads into a system the next team can walk into. An agent reads it every morning. The orange sweep is the read.
drag to turnyou
10 agents spun up, each asking the memory (brain) from a different point of view0 of 10 done
synthesis
answer readyExport PDFExport HTML
reading the memory (brain), every morning, on its own
- next best action
- Action 1Three directorates show the same grade 14 range-exception pattern as last cycle. Propose a pre-approved exception rule so those requests stop waiting.to humanneeds the comp lead's decision
- Action 2Survey submission window opens in 11 days. Pre-fill the file from last cycle's mappings and flag the 18 new jobs.subagent runningsubmission agent, 74%
- Action 3212 benchmark matches drifted since the last refresh. Re-matched, rationale logged against each job.analysis completeexecuted; 3 low-confidence matches held for review
- recommendations
- #1Route in-range off-cycle requests to automatic approval.
Cuts the request cycle from 6 days to 2, based on the last 200 requests in memory.
- #2Attach a plain-language "why" to every decision shared with HR business partners.
Ends the black box. HRBPs asked "why" on 60% of flags this quarter.
- #3Move two people's worth of transactional time into advisory work next quarter.
The freed capacity shows up in the memory as unclaimed hours.
Section 4, one workbench
One workbench brings it together: every team member and every AI coworker in the same workspace.
One thread with the main agent on top. Below it, the approvals your AI coworkers queued overnight, with the data already pulled from Workday and the other systems. Guardrails apply to every agent the same way. Every interaction enriches the memory (brain).
compensation workbench
deskapprovalstracksmemory (brain)
17 AI coworkersguardrails ontue 07:42
thread with the main agentcomp workbench agent
Why was the adjustment for employee 48213 flagged?
It lands 6% above the range maximum. Third adjustment in 24 months; the last two ratings were "meets". Market sits at P45 for the role, pulled from Workday and the survey cut. Three options: hold at range max, approve with a 12-month review, or re-level to grade 14. Each has been used in this directorate before; the memory (brain) shows which one stuck.
Hold at range max with the 12-month review. Log the rationale and tell the HRBP.
Done. Rationale logged against 48213, HRBP notified in the thread, and a rule candidate added to the memory (brain): third adjustment in 24 months with two "meets" ratings holds at range max.
Approvals waiting for you4
AI-led
benchmark match agentovernight, 02:10guardrails passed
212 jobs re-matched to the new survey cuts. 209 above the confidence line. 3 need your call before the file goes out.
Review 3Approve 209See rationale
AI-assisted
merit planning agentprepared 06:40guardrails passed
Budget allocation draft by org, three scenarios. Scenario B exceeds the pool by 0.3%: approve with a cap, or send it back.
Open draftApprove with capSend back
AI-led
exception review agent07:05approval above $5K
Two range exceptions above $5K, pulled from Workday with market P45 attached and the precedent from last cycle. Nothing to look up; both are ready for a decision.
Approve bothReview
autonomous
survey submission agentready, 07:30guardrails passed
Submission file assembled from last cycle's mappings. 18 new job mappings proposed from Lena O.'s cut. Approve to send.
Approve and sendEdit mappings
memory (brain), today
+14decisions
+3rules
+41rationales
learned this week
- Grade 14 exception pattern repeats in two directoratesfrom exception review, confirmed by Maya R.
- HRBPs ask "why" on 60% of flagsfrom the thread log
- Survey cut mapping for 18 new jobsfrom Lena O., reused by survey submission
human-led
I start the conversation
I go to work and open a chat. The agent answers from the memory, with the history and the rationale behind it.
AI-assisted
I run it, the work is prepared
The analysis, the draft and the exceptions are already next to me when I open the process.
AI-led
The AI took the initiative
Agents worked overnight and are waiting for me to approve or give feedback.
autonomous
It does the job, I stay informed
Hardened, guardrailed, tested, and living among the humans. It reports; it does not ask.
Why workbench?chat system vs workbench
| Capability | Chat systemChatGPT, Copilot, Claude, Cowork | Workbenchwhat you see above |
|---|---|---|
| Switching systemsthe main difference | You leave the chat to pull data from Workday and other internal systems, then paste it back and explain it. | Integrations let the agents pull the data to you. You stay on the task instead of switching interfaces. |
| Collaboration | One person, one thread. | Coworkers and agents in the same thread, with the people who own the process. |
| Publishing AI coworkers | Personal assistants, one per user. | Publish an AI coworker once; the whole team interacts with it. |
| Guardrails | Per-user settings, if any. | Enforced as a standard across every agent's capabilities. |
| Memory (brain) | Stays inside the vendor's chat, per user, and leaves when the subscription does. | Captures the cognitive loop into an organizational memory (brain) that lives outside any chat system. |
| Customization | A fixed experience. | Customize the experience, the output and the reports to the team. |
| Who starts the work | You, every time. | Agents work overnight and queue approvals; you still start whatever you want. |
The ultimate goal
In the future, most business decisions will be made by agents.
They should not make those decisions on LLM reasoning alone. An agent needs the right context to take the most appropriate decision. Agents use the LLM for reasoning. They use the memory (brain) to apply that reasoning with the right context. That is why the organizational memory (brain) is the goal.
Scroll sideways to follow the flow.
Same question, two answers
Should we approve the $6,200 off-cycle adjustment for employee 48213?
An agent on LLM reasoning alone
"An adjustment above the range maximum is usually declined unless there is a documented retention risk. You may want to check the compensation policy and consult HR before approving."
- correct in general
- knows nothing about this person, this directorate, or last cycle
- hands the decision back to you
The same agent with the memory (brain)
"Hold at range max with a 12-month review. This is the third adjustment in 24 months, the last two ratings were 'meets', market sits at P45, and this directorate chose this option in 31 of 38 similar cases. Rationale logged, HRBP notified."
- the same reasoning, applied to what is true here
- cites the history, the rule, and the precedent
- decides, and the decision teaches the memory (brain)
the cumulative score=the memory (brain) your agents will decide with
Every track adds to the score and to the memory (brain). One problem at a time is how you build the context that turns agent reasoning into the right decisions.
One question to leave with