Unique Research · Insight / Feature · January 10, 2024 · Author (CN): Tan Qiuping · Guests: Zhou Jian, Founder & CEO, Lanma Tech (澜码科技); Liu Yuchen, Founder & CEO, Tyrion.AI (意言科技)
In a July 2023 blog post, former a16z partner Benedict Evans wrote that almost everyone in tech agrees generative AI, large language models, and ChatGPT mark a generational shift in what software can do — and in how far automation can reach. Even where people disagree about LLMs, they converge on one point: more automation is coming, new kinds of automation will appear, and that will touch jobs and how organizations manage people.
Canadian political economist Harold Innis argued in The Bias of Communication that every new technical invention creates a new service environment and a large-scale recombination of social experience — and that “recombination of the sense of life” underpins social continuity. Environments we invent almost without noticing place heavy demands on consciousness and knowing.
If you have felt the force of LLM-driven AI Agents remaking process automation, Evans and Innis land as lived experience. How will Agents reshape enterprise process automation in concrete terms? Unique Research spoke with two representative China Agent founders — Zhou Jian (Lanma Tech) and Liu Yuchen (Tyrion.AI) — about the value Agents already create when they automate business flows.
Classical process automation vs LLM-era Agent automation
Zhou argues LLMs mainly unlock three capabilities. First: extremely cost-effective language understanding — decisive when enterprises build new automation apps or Agents — which also makes conversational UI real. Machines begin to grasp question meaning and fill subjects and phrases from short descriptions. Where people once adapted to machines, machines now adapt to people.
Second: strong comprehension and summarization that can integrate more sources, including semi-structured documents — sharply raising document-information delivery and task-extraction efficiency in commercial settings.
Third: logical reasoning — e.g. text-to-SQL, or picking the right API from a hundred candidates from text alone.
Those abilities help automate and intelligentize downstream processes and multi-activity work. Classical RPA mainly automated operations — intelligently decomposing steps for point fixes. LLMs can start from the whole flow. Every SOP has dimensions across data, documents, and application process; on LLMs, data and documents should be priority: text-to-SQL, text-driven BI tools, document extraction, document compare and statistics.
Many top internet firms already ran AI-driven process automation in scenes — but ROI was often poor. In the LLM era, firms can use higher cost-performance, more SaaS-ready, smarter infrastructure to drive process automation.
A classic pain: everyone joins many WeChat Work and Feishu groups; information scatters. When a project owner changes or a client raises a key need, staff must manually scroll history. With AI, firms can automate collation and analysis. Separately, AI can amplify expert leverage: scarce senior capacity once blocked many tasks; experts can encode experience into Agents that strengthen frontline staff at each step — lifting process quality.
Zhou also stresses AI as a Connected role. A process spans many roles; a task or activity belongs to one. Firms once used BPM or workflow engines to move work across roles. Agents can design automation that goes beyond — even replaces — those engines. Inside Lanma, developers have explored Agents for auto-subscribe, booking, monitoring project/customer change and progress.
Gartner’s Top Strategic Technology Trends for 2024 lists the Augmented-Connected Workforce (ACWF): a strategy to optimize employee value with intelligent apps and workforce analytics that provide daily context and guidance for experience, wellbeing, and skill growth. Workplace automation and AI make smarter work possible — and demand that employees manage more complex problems.
How Agents realize process automation
1. Foundation: split process & key information; digitize expert knowledge
Automation first needs decomposing and organizing information passed across activities. Zhou cites a Lanma-served recruiting firm that set precise commission splits from consultant to delivery on every order — a layered standard that motivated every role to log business information to prove contribution. Agents can then extract rich signals for process remakes.
In short: digitize expert knowledge by clarifying flows; with clear business dictionaries, rules, and cleaned process history, use methods that let process data settle — then Agents can substitute with value.
Enterprises long automated to “cut cost, raise efficiency.” Zhou’s read: Agents’ core value is raising work quality for business staff. Competitive edge often equals top-expert level — but expert dependence is risk. Agents can clone expert “twins”; even ~70% of expert level is valuable, and GPUs can scale response capacity. So Agents today are mainly about revenue lift and compliance — not only cost cut.
Example: recruiting firms live on continuous efficiency gains. A software-outsourcing recruiter org with 100 recruiters might see most hire ~1 person/month while the best hire ~10. Encode the top recruiter into an Agent so the rest average 4–6 (or more) hires/month — and revenue rises. That is the problem Lanma is solving, and where Agents show clear value.
2. Execution: Golden SOP and multi-Agent collaboration
Tyrion.AI builds Tyrion — an LLM-driven “middle-layer” Agent framework and platform. Agents built on Tyrion encode expert knowledge, embed into workflows, and collaborate with employees on complex tasks.
Liu says expert knowledge and SOPs are key operating assets. Before Agent build, Tyrion helps clients clarify demand: which processes can consolidate into reusable Golden SOPs that guarantee landing effect. With LLMs, SOP execution also becomes more flexible. Excel, calculators, and classic software once replaced only fixed, logical tasks — not human judgment or soft communication. Expense claims still needed verbal confirmation beyond system submit; ERP owned “submit,” not the conversation. AI can automate some of those flexible communication moves.
Real enterprise work is multi-person. Replacing some people with Agents means human↔multi-Agent and multi-human↔multi-Agent interaction: when to insert humans, when Agents, which part Agents own, how machines collaborate with staff. In plain terms: remake formerly human-only flows into machine + LLM + Agent + human. Effective human–Agent collaboration is the core challenge.
Tyrion is exploring new interaction that binds multiple Agents with multiple humans — and lifting Agent depth of thought and intent understanding for smoother co-work. If every workflow is human-only, it feels ordinary; no firm yet routinely runs multi-Agent participation at scale — that is what Tyrion is pushing.
3. Core remake: digitize employee behavior
Extracting industry know-how scattered in individual experience has long been painful. Firms often tried “knowledge centers” or mid-platforms. Zhou: knowledge is partly codifiable, partly experiential. Experiential knowledge can now be managed by digitizing employee behavior via Agent-automated flows, then structured with ML or LLMs.
Liu grants that remaking process and human–machine collaboration mindsets is hard and long — gradual learning and acceptance. But as AI digitization advances, enterprises will converge there. Digital employees will be more reliable than humans on specific tasks — a main substitute force. Industrial automation took one slice of work; software took another white-collar slice; Agents drive a third stage of process automation, replacing more repetitive labor — not total replacement, but organic collaboration.
In the industrial age people asked which line tasks machines should own; early software asked which workflow slices software should own; the LLM/Agent era asks the same. The challenge: few precedents — demand discovery must scale Agent use, and that is still early. Embedding software into processes once faced friction and upgrades, but forms were simpler. Agents act more complexly and involve non-software judgment. Layers: enterprise software at bottom; RPA on software ops (optional — other tools can replace); Agents above; human interaction on top.
Agent-driven automation will remake the ToB software stack
The hard question: LLM- and Agent-driven process upgrades that remake white-collar work will disrupt enterprise organization and ToB software. Zhou and Liu offer views.
Max Weber’s ideal bureaucracy — clear hierarchy, duty division, authority, formal procedure — was long the template for organized efficiency. Zhou argues that in the LLM era this static hierarchy will shift. Alibaba-style mid-platform orgs will gradually be replaced by Agents. Future structure may become large front office + large back office; mid-layer information pass and summarization fall to Agents that brief department and mid-platform resource owners for decisions — hierarchy toward more agile automation.
Zhou also sees classical B-side software as long bottlenecked: DevOps, AI 1.0, and many techs were one-shot breakthroughs; RPA mostly solved app-layer final clicks. LLM language understanding and reasoning are general breakthroughs usable across scenes — and ToB commercial logic shifts with them.
Past ToB monetization: find the right buyers and decision-makers by industry/function. China’s IT budgets concentrated in SOEs, banks/finance, and buyers such as CFOs, CHOs, HR, marketing. Now Agents can substitute on-role — cutting headcount budget directly — a different logic from classic ToB vendors, opening new opportunities and a dimensional strike on incumbents. Lanma currently maps which domestic roles have high reuse/replicability, digitizes expert knowledge via Agents, and pursues human–machine fusion.
Zhou adds: training vertical industry LLMs may not be the right path. Models compress knowledge. Same-industry different roles overlap a lot; same-role across industries overlaps even more; different-industry different-roles overlap little. A bank president and an insurer CEO overlap high; a bank Java engineer and an insurer Java engineer overlap high; a Java engineer and a bank president overlap very low. That creates room for Agents to remake org process — and role-angle entry is more replicable. Agents ultimately compete for HR budget: substitute roles and lift enterprise return.
In 1913 Henry Ford, drawing on Taylor’s standardized work, built the first moving assembly line — high efficiency, low cost, high wages, high profit — then ERP integrated processes and departments. Knowledge-worker flows still lack standardization; there is no “ERP for knowledge trades” — a structural opportunity. Zhou: upper managers need standardized interfaces from staff whose work is soft and personal. Agents resolve that tension by standardizing employee process — foundation for remaking ToB software form.
Liu lands nearby: LLMs and Agents open disruptive opportunity in enterprise software and unlock latent demand — so AI-native companies can build Salesforce- or SAP-scale giants. Salesforce, Adobe, and SAP shipped many AI apps last year — mostly bolted onto existing feature points. Real leverage means rethinking software AI-natively: which scattered, formerly human-only flows can machines own now?
As Agents evolve, many legacy processes and tools face full disruption. Agents seep like sand into every enterprise “seam,” remaking end-to-end flows and scenes.
Agent as a Service — a productivity engine for everyone
At OpenAI’s first developer day (Nov 2023), GPTs, the GPT Store, and the Assistants API opened the Agent commercial-ecosystem curtain. As more players, clearer architectures, and stronger base models arrive, Agent as a Service as a productivity mode is no longer only a vision. Zhou’s hopeful picture: domain experts put knowledge on cloud services via dialogue (or other forms) and sell their Agents for larger influence.
Yet iterating models, shifting how people work and think, and finding true best practice will not be easy — possibly with pain and conflict. Before that, Zhou suggests clarifying values: is the Agent process-automation revolution meant to help the “achievement subject” break limits and escape forced freedom / over-self-exploitation — or to reinforce the achievement-society paradigm and its compulsory structures?
1 CN image; attach on ingest. Author (CN): Tan Qiuping; edit: Qiuping.
— END —