Unique Research · Interview · August 29, 2024
In AI tech, RealAI (实在智能) has carved out a strong market position with its self-developed AGI large model and hyperautomation stack, and a steady push in human–machine collaboration.
Since entering RPA, RealAI has shipped three generations of RPA products through successive tech upgrades. The third generation further combines large models with RPA, pushing the industry toward the Agent era and a higher level of human–machine interaction.
In August 2023 RealAI publicly released its self-developed TARS large language model and the industry-leading TARS-RPA-Agent — RealAI’s RPA Agent. Natural conversational interaction plus hyperautomated execution further simplifies building digital workers and makes the tech easier for a wider user base to adopt.
Recently, RealAI Vice President Chen Wei joined Unique Research to share the company’s practice and results in Agents.
Interview transcript below.
RealAI’s Agent landing practice
Q: RealAI started with RPA tooling and has shipped three generations. You’ve now moved further into Agents. What were the key milestones — and what mature practice do you have in Agents today?
Chen Wei: Entering Agents was grounded in a long accumulation in AI. From day one RealAI focused on AI R&D. Along the way we saw that RPA, as an automation tool, gives customers a relatively easy on-ramp to land automated digital scenes quickly.
Intelligence, by contrast, asks more of the data foundation. So RPA became a preferred path for customers in early digital transformation who needed automation projects to land fast.
Our first-generation RPA shipped around 2020, aimed at professional developers. It abstracted PC operations — Excel, browsing, Word, and so on — into action components. Drag and drop those components to edit an automation flow.
That generation accumulated about 400 preset components and thousands of rich business components. It simplified development, but still had a learning cost — especially hard for business users.
To lower the bar, generation two brought in more AI: intelligent component recommendations based on smart screen semantic understanding. Point your mouse at an object on screen, and the system suggests the best component. Users no longer hunt through thousands of components; they develop and record flows from AI suggestions mid-operation.
If gen-one was “drag-and-drop,” gen-two is “point-and-pick” — usable without deep tool training. Our patented smart screen semantic understanding lets digital workers dynamically recognize desktop elements the way a human eye finds input boxes and buttons.
On that base we folded in LLM generation for generation three — the Agent digital worker. The model understands tasks and flows and turns a task description into concrete computer steps.
Example: “Fetch the latest 20 news items on company X.” The system decomposes that into: open browser, search keywords, click news, collect items. As the PC runs those steps it calls RPA components (“open page”), uses screen semantics to find the input box, and the LLM generates components to type the query and click search.
That yields smarter, more automated execution — higher efficiency and a better user experience.
Q: In what scenes do customers typically use this?
Chen Wei: A typical first landing was carrier customer service. Pure Q&A can be handled by smart CS pulling answers from a knowledge base. But take a user who dials 10086 to change broadband: after confirming details conversationally, traditional process still requires an agent to type the change into a back-office ticket system and hand it to frontline staff. Ticket ops used to be human-only; an LLM-powered RPA-Agent can now enter those tickets automatically.
Through dialogue the Agent captures a true, accurate, complete request and replaces human ticket entry — so the full broadband change can run as a digital worker with no human in the loop.
A second class of scenes is learning and operating across systems. Train an Agent on Yonyou or Kingdee finance; if the customer runs Inspur, an Agent that already understands finance ops and process can pick up Inspur quickly — vouchers, reports, and so on — and keep refining the best path for later runs.
So many system ops will eventually be language-instructed: digital workers execute; humans plan and check results. We call this system-ops takeover via LLM + RPA-Agent.
Beyond that, RealAI Agents support retrieval-augmented generation: no pre-training required — hang an enterprise knowledge base for professional Q&A, like an open-book exam for the LLM. Traditional smart CS turns knowledge into Q&A pairs for training — costly and hard to update. With an external knowledge base, upload new material and answer immediately. That architecture is especially useful for internal policy Q&A — expense rules, travel subsidies by location — low training cost, strong effect.
The same idea flips into review: an Agent can check expense claims against finance policy in place of a human reviewer.
Finally, Agents combine with analytics tools to pull data from databases or a data middle platform on request and build charts. Ask for January–July user distribution; the Agent produces stacked or line charts as fit, and the LLM can comment on trends or issues — more timely, personalized analysis.
Q: Extending the finance example — could we train digital workers with different skills, package them, and offer them ready-made to enterprises?
Chen Wei: Exactly. Most vendors still sell traditional tool software. The trend is toward selling trained digital workers that already hold the skills for finance or other process work — effectively a new kind of manpower outsourcing that supplies real digital employees.
That means a virtual team covering 80%+ of basic work at far lower cost than humans, raising throughput while cutting spend.
Q: You mentioned innovation on external knowledge bases. How do you do it, and what’s the edge?
Chen Wei: Several layers. Most external KB systems sit on RAG; we improve on that. Knowledge construction needs vectorization and preprocessing. Language models mostly handle text, so for charts and images we use IDP (intelligent document processing), OCR, and table-recognition plugins for accurate extraction. We also clean headers, footers, and special symbols that confuse understanding — better structure and completeness, higher model efficiency.
Second, in domains like finance and contract review we’ve built rich review-point and rule knowledge — nearly out-of-the-box in common scenes, better value for customers, and a market edge.
We also add an FAQ precipitation path for high-frequency questions: analyze ask frequency and content, distill standard answers with the LLM into a high-frequency FAQ store. Repeat questions hit FAQ first instead of regenerating — lower compute load, and a concrete view of what users care about and how well it’s handled.
Q: The solution looks mature — landing approach included.
Chen Wei: That comes from experience and algorithm depth. Among peers of similar size we have nearly 30 algorithm staff — unusual competitiveness. Most RPA vendors have few or no algorithm people or patents; our reserve is high.
Deep NLP and CV let us follow the 2021 LLM wave quickly with our own vertical model, tightly combined with existing RPA, IDP, and BI — so existing customers get an LLM upgrade path.
Many pure LLM vendors still hunt for scenes from zero. We can apply LLMs directly to current RPA/IDP/BI customers so they feel the upside immediately — lower conversation friction, easier new-logo motion.
Q: With Agents, does RPA’s room shrink? And for customers already on classic RPA, how do you upgrade to Agents?
Chen Wei: We’ve seen OS shifts from DOS CLI to Windows GUI that democratized computing. Bill Gates predicted interaction would move to Agents. As interfaces become APIs, classic RPA demand may fall because large models can call API-ized apps directly.
That shift is gradual. Screens and screen-based ops will remain for the foreseeable future. Wiring APIs into a large model is relatively easy — any firm can do it. Letting a large model drive RPA on the desktop is a harder technical barrier.
First, we already have a mature, leading RPA product validated across thousands of customers.
Second, mainstream LLM vendors still don’t show strong image-level desktop element recognition — limiting desktop control. Our core edge is screen semantic recognition — multimodal vision to locate and control desktop elements.
Third, our model is pre-trained and fine-tuned on process knowledge, so on process problems it can outperform much larger general models even if it is smaller than GPT-4-class systems.
Until full API-ization, RPA-driven desktop ops may remain the only practical substitute. As we accumulate business logic and knowledge, our model can seamlessly take over tasks RPA once ran. Pure API Agent vendors without enough business experience may still struggle on business understanding even if they can call APIs.
Looking ahead, we expect RPA to move from a front-office tool to a system substrate. We’re exploring with system vendors embedding RPA so systems expose underlying RPA control to large models, combined with API control — stable desktop control with good transitional results. That’s our long view on Agents and RPA.
How Agent applications are evolving
Q: How do you read the domestic Agent landscape?
Chen Wei: Agents have moved through clear stages. Before 2021, Agents were mostly classic chatbots calling pre-built process APIs — mechanical, fixed, low technical difficulty, able only to run services that matched the knowledge base.
From 2021 to 2024, with large models, many vendors tried to land Agents. Mainstream AI-Agent tools often configure API or model capability workflows on LangChain. We tend to call those reactive agents: many APIs, execute what the user asks — but capability is capped by what was integrated at design time; limited innovation or self-iteration.
What we’re building gives Agents two core abilities: learn human business, and control the computer. With business knowledge, an Agent can plan and execute from a human plan or task. We call that an autonomous agent — it knows how the business runs and can plan computer ops to enact it. If it learned one finance system, it can, via desktop recognition and self-learning, pick up a never-seen finance system and run similar ops — theoretically unbounded learning.
The next stage is collaborative intelligence: domain-expert agents that learned separately start coordinating with each other for broader, more efficient workflows.
Q: You stressed “capability boundaries.” Do some customers over-expect AI Agents — a big gap to reality? What challenges show up when deploying?
Chen Wei: Early in the LLM wave — especially last year — many top customers poured resources into advanced models with high expectations and mixed results. Our expectation management leans toward upgrading on top of products customers already value, not dumping entirely new tech on them.
Customers already feel value from classic RPA or IDP. Combining LLMs with those foundations lifts results in a controllable way. Training an expert model used to take heavy tech and headcount; with an LLM, a single spoken command can deliver a clear lift.
LLMs are powerful but not omnipotent. In RAG they don’t always retrieve every piece of knowledge perfectly — so we help customers restructure knowledge, improve presentation or output shape, sometimes add labeling. Versus classic model training, that’s already simpler: more like talking business logic and turning it into prompts. The simplest prompt is the user’s own sentence — every customer can join.
Customers can also use our Agent tooling themselves, without us in the loop, to tune or invent for specific scenes — more flexibility.
On pricing we aim for a level that encourages broad adoption: win share with technical lead, build trust, create longer-term optionality.
Early large-model releases didn’t always meet expectations; with RAG and Agent architectures, customers are already feeling clear product lifts.
In RPA practice, users face many specific but low-frequency long-tail scenes. Custom-building each one adds cost that doesn’t pencil out — especially for business users who need the automation but can’t develop it. An RPA-Agent helps: describe the flow orally (more detail the first time); the digital worker runs it; if blocked, the Agent asks for help or whether to adjust steps. That interaction learns and stores process knowledge so the next run is much faster — solving long-tail construction.
Some scenes classic RPA couldn’t handle well — e.g. dynamic planning. Customer-service tickets need re-planning entry fields by request and typing into non-fixed UIs. We built a hospital intelligent dispatch system where Agents schedule caregivers by real-time location for optimal routing — much higher service efficiency and response. In data monitoring, Agents can detect anomalies and analyze — again beyond classic RPA.
Q: With so many Agents in the market, can they eventually form one automated workflow — and on what timeline?
Chen Wei: Two main challenges. First, information sharing: firms that hold valuable data and knowledge often won’t fully open them; large platforms tighten protection. In competitive markets that’s common. Full integration is a long game — technical, commercial, and privacy — dependent on cooperation mechanisms and sharing standards.
Second, by industry: distinctive specialist agents will excel in their domains. They may never share one industry-wide standard, but leading agents will matter most. By the 80/20 rule, ~20% of leading agents may create ~80% of productivity. Head agents will drive progress and standards; as standards settle, collaboration rises into more coordinated automated workflows.
Q: In the LLM and Agent era, how do enterprise data accumulation, governance, and management change?
Chen Wei: We’ve proposed an “enterprise brain” that integrates large models with internal operating architecture to upend classic information systems — Agents take over the OS layer and remake the stack.
Classic IS was designed for humans to process information, but often raised operating cost and lowered efficiency because of complex data dimensions and skill needs. LLMs change that. An enterprise brain gives every employee an enterprise-scoped expert knowledge base. Business and data knowledge become systematic knowledge operations. Future competitiveness hinges on mastering and applying industry expert knowledge.
The brain puts large knowledge resources to efficient use. Future IS gets simpler: functional UIs collapse into one conversational entry. Classic expense systems are heavy with rules and config; now an employee uploads receipts in chat and the model matches records and runs review.
Q: Agent capability is largely capped by the underlying model. How do you see that relationship? If LLM progress slows, does Agent commercialization suffer?
Chen Wei: LLMs are moving in two directions. One is ultra-large models — GPT-4o, GPT-5, and peers — like powerful engines for broad or high-value general scenes. For bounded business value we often prefer smaller models that are more efficient and cheaper on specific tasks; we iterate and train them inside the enterprise. They won’t grow as fast as frontier models, but fit concrete work better.
So we focus less on forcibly fusing maximum model power and more on decoupling — giving customers choice. Use our model or another; we handle deploy, serve, train, and tune. Multiple large and domain models collaborate through our Agent workflows to hit business goals.
Decoupling also means enterprises can plug in the newest models while internal knowledge bases and external tool capabilities keep accumulating — often more important to the business than raw base-model gains, especially once cost and ops are counted.
Only for very special capability needs or rich proprietary data would we advise training a vertical industry model — which, once trained, can serve the customer and the wider industry.
Interviewer: Unique Research Editor-in-Chief Qiuping
Editor: Qiuping
Registration / follow-up channel: see original Chinese post.
— END —