Appian Claims 80% Token Cuts by Putting AI Agents Inside Processes
Appian expanded its AI-agent control plane and launched Ask Appian in London, with Raiffeisen reporting 80% cuts in token use and processing time from a hybrid agent-and-rules approach.

Why it matters
- Appian unveiled Ask Appian and an expanded AI-agent control plane at Appian Europe in London on November 17, 2026.
- Raiffeisen Bank International reported an 80% reduction in token consumption and processing time from a hybrid agent-and-rules approach in testing.
- A study by Appian and Harvard Business Review Analytic Services found 92% of respondents say agents need rules-based guardrails; only 48% have them.
- Cross-platform tracing, A2A protocol support and third-party agent measurement remain on Appian's roadmap, not generally available.
Raiffeisen Bank International says a hybrid agent-and-rules approach cut token consumption and processing time by 80% in testing, and Appian is building its entire AI-agent strategy around that number. At Appian Europe in London on November 17, 2026, the company expanded its AI-agent control plane and unveiled Ask Appian, a plain-language interface that dispatches agents across connected enterprise systems.
The announcement matters because enterprises are struggling to govern AI agents (autonomous software programs that perform multi-step tasks) once those agents move beyond demos. Appian's answer is structural rather than technical: put the agents inside deterministic business processes, where the process — not the agent — decides what happens next.
What does the expanded control plane actually do?
The control plane registers both Appian-built and third-party agents in a single management view. Administrators can monitor agent activity and cost, restrict which models agents may use, track policy compliance and generate alerts when something drifts.
An optimization layer sits on top. It runs evaluations (structured tests that measure agent accuracy) and connects each agent to the specific cases and process steps it handled. Business key performance indicators — cost per case, cycle time — supply the operating context, so a manager can see not just whether an agent behaved, but whether it produced a cheaper or faster outcome.
Sanat Joshi, Appian's executive vice president of product management, said permissions, guardrails and model access cannot establish whether an agent is doing useful work. "Appian puts business process at the center," he said. The claim is defensible as far as it goes; the caveat is that this measurement currently works only for agents running inside Appian's own platform.
What is Ask Appian?
Ask Appian is a separate interface where employees request work in plain language. It pulls relevant context from systems connected through Appian's data fabric (the company's unified layer over disparate enterprise data sources) and dispatches agents to complete the task.
The stated goal is modest and concrete: spare employees from hopping between underlying applications to assemble information. A worker types a request; agents gather the context and execute.
Why does Appian insist agents live inside processes?
Appian's architecture assigns every step in a business process to an agent, a person, a system or a rule. Rules handle predictable work. Agents receive guardrails and shared context for steps requiring judgment. If an agent fails or is stopped mid-task, the work falls back to a person, a rule or another agent, so the case continues rather than stalls.
A November 16, 2026 Appian blog post spells out the trade-off with unusual candor. Process-led steps are consistent, repeatable and auditable; agent-led steps handle open-ended work but add cost and latency, and outcomes can vary case by case. In one worked example, an agent assessing a loan hands debt calculations to deterministic formulas while a human loan officer makes the final decision.
The takeaway in plain words: Appian treats agents as replaceable components inside a controlled workflow, not as autonomous actors.
What works today, and what is still on the roadmap?
The documentation is explicit about the boundary, which is to Appian's credit.
Available now:
- Registering Appian and third-party agents as steps in Appian processes, through MCP (Model Context Protocol, an open standard for connecting AI systems to external tools and data) or direct integration, without moving the agent into Appian
- Controls, orchestration and process intelligence within the control plane
- Measurement of Appian agents against process KPIs
- An early-adopter program granting access before general availability
Planned, not shipped:
- Support for the A2A protocol (Agent-to-Agent, a protocol for agents built on different platforms to communicate)
- A registry encompassing third-party agents
- Tracing for agents running outside Appian
- Estimated cost per case and per step for Appian and third-party agents
- Before-and-after comparisons with manual work
- Agent-improvement recommendations
Appian said additional capabilities would start arriving during 2026. The company itself notes that planned features and timing can change, and that purchasing decisions should rest on generally available functionality. That disclaimer deserves more attention than press-release readers usually give it.
What did Raiffeisen actually test?
Raiffeisen Bank International intends to automate payment inquiries with Appian. Claus Mitterlehner, its head of smart automation, said the team initially gave most tasks to AI agents. "After testing it, the team changed course and moved the predictable steps to deterministic automation, keeping agents for the steps that need judgment," he said.
The 80% reduction in token consumption and processing time came from that correction — not from agents doing everything, but from agents doing less. Raiffeisen did not publish the absolute figures behind the percentage, the workload size, or the testing duration; the claim rests on the bank's own reporting of internal tests.
This pattern illustrates the control plane's broader optimization loop. Evaluations and process KPIs identify how agents perform, and repeated agent steps become deterministic workflows or business rules. The intended end state is a process where agent spending concentrates on judgment-based work rather than predictable execution.
What do buyers say they need?
A study by Appian and Harvard Business Review Analytic Services, which surveyed AI decision-makers on modernization and workflow integration, found a gap between intention and implementation: 92% of respondents said agents need rules-based guardrails to be safe and effective, while only 48% had such rules in place.
The study was commissioned by Appian, which is an obvious caveat, but the 44-point gap aligns with the company's pitch — enterprises want controlled agents and mostly do not have them yet.
What is still unanswered?
Appian's control plane still lacks cross-platform tracing and measurement for agents built outside its platform. Until the A2A support, third-party registry and external tracing ship, the unified management view is unified only for agents that play by Appian's rules. The 92%-guardrails statistic suggests demand; the 48%-implementation statistic suggests most enterprises have not solved this anywhere yet.
The useful test of Wednesday's announcement is whether process evidence can show where agents improve cost, speed or accuracy — and where repeated work should become rules. Raiffeisen supplies the one concrete data point. Whether Appian can extend that analysis across agents built elsewhere will determine whether the control plane becomes genuine infrastructure or a well-instrumented walled garden.
Source: unite.ai
More from James Calloway
Show full bio
News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.
722 articles
Related articles
- Uptiq launches GovernAI to govern AI agents in banking at no extra cost
- Australian firm opens the books on five months of AI agent controls
- Airlock Digital adds action-level controls for Codex and Claude
- 83% of Firms Run Unapproved AI Coding Agents, BlueFlag Research Finds
- BNY Runs 125+ AI Use Cases in Production on OpenAI-Powered Platform