The question organisations were asking two years ago was “can AI agents actually do this?” The question most are asking now is “what does it take to operate them once they’re running?”
The technology has moved faster than the operational frameworks around it. G2’s August 2025 enterprise survey found that 57% of companies already have AI agents in production — up from a fraction of that a year earlier. 80% of enterprises that deployed agents report measurable ROI, with U.S. enterprises reporting average returns of 192%. The adoption curve is steep; the operational learning curve is steeper.
What does production actually look like? Here are two builds from our delivery work — what we set out to solve, how we built it, and what the results revealed about operating autonomous agents at scale.
Case One: Replacing 70% of the Bid Team’s Admin Work
A European B2B services company had a problem common in competitive tendering environments: their most expensive people — bid managers and pricing analysts — were spending most of their time on work that did not require their judgment. Document review, deadline extraction, certification matching, portal tracking. The estimate was that roughly 70% of these senior roles was administrative.
The bottleneck was not the quality of their bids. It was the sheer volume of monitoring and preparation work that had to happen before any strategic thinking could begin.
The system we built uses a four-agent architecture:
- A scraping agent monitoring tender portals on a defined cadence, flagging new opportunities against the client’s eligibility criteria
- A parsing agent processing PDF documents — specifications, requirements, terms — into structured data
- An extraction agent identifying requirements, deadlines, certification needs, and scoring criteria from parsed documents
- An analysis agent assembling structured outputs for the bid team: a prioritised brief on each opportunity with extracted requirements pre-matched against existing capabilities
The technical stack: FastAPI microservices for independent agent scaling, PostgreSQL for state management, OpenAI for reasoning via LangChain, Docker for deployment, and rate limiting controls on portal interactions.
The results from the first six months of production: submission volume increased 400%, compliance accuracy approached 100%, and the bid team’s time — previously 70% administrative — shifted toward pricing strategy and competitive analysis. The headcount did not change. The output did.
The lesson that shaped everything that followed: production agent value rarely shows up as “AI replaces person.” It shows up as “AI clears the runway so the person can do the work you were actually paying them for.”
The full build detail is in the AI-powered Tender Optimization case study.
Case Two: Capturing Revenue That Used to Evaporate Overnight
A European travel agency selling last-minute deals and regional tours had a different constraint: a content production bottleneck that was costing them bookable revenue.
Two marketing staff were generating 5–7 social media posts per day. Business requirements called for 15–20. The gap was widest at the moments that mattered most commercially: evening and weekend launches for time-sensitive departures with available seats. Those opportunities were generating no promotion because no one was working when they needed to go out.
The four-agent system:
- A planner agent reading the tour management system continuously, prioritising departure timelines, seat availability, and margin data to determine what needs promotion and when
- A copy and asset agent generating channel-specific captions — format and tone adapted per platform — and retrieving approved imagery from the asset library
- A compliance guard verifying brand voice, pricing accuracy, and required disclaimers before anything is published
- A scheduler publishing via platform APIs, applying UTM tracking tags, and feeding engagement metrics back to the planner to inform future prioritisation
The technology stack: LangGraph for orchestration (the branching logic between agents required explicit state management), Pinecone for the knowledge base indexed with brand guidelines and historical high-performers, secure MCP servers integrated with the tour management system on read-only access.
Production results over six months: content output rose from 5–7 to 12–14 posts daily without additional headcount. Seat fill on promoted tours improved 14%. Marketing staff time spent on publishing dropped 58%. The previously lost revenue category — urgent overnight deals launched without promotion — achieved 100% capture.
The full build is documented in the Travel Agency Content Automation case study.
What Is Actually Different About Production in 2026
Both builds taught the same operational lessons. They are not unique to these clients.
Orchestration framework selection has long-tail consequences. The tender build used LangChain for a sequential pipeline where that approach was appropriate. The travel agency build required LangGraph because the branching logic between agents — planner routing to copy agent, compliance guard routing back for revision, scheduler routing engagement data back to planner — needed explicit state management. Framework selection is not a technology preference question; it is an architecture decision that determines how maintainable the system is when it inevitably needs to change. We have a full breakdown of the current framework landscape in our AI agent frameworks article.
Knowledge base architecture is more important than embedding model selection. Both systems use vector storage for retrieval. In both cases, the critical decisions were about indexing structure — how brand guidelines sit alongside historical high-performers, how product data is embedded with margin context, how regulatory clauses are tagged for retrieval — rather than which embedding model to use. Teams that spend weeks on embedding model comparisons and days on indexing architecture typically regret the prioritisation.
Governance cannot be retrofitted post-deployment. Essential governance elements — read-only integrations to authoritative data sources, schema validation on all agent outputs, least-privilege authentication keys, comprehensive audit trails, human-in-the-loop approval gates during initial deployment — need to be designed in from the start. In the tender build, the compliance accuracy result depended entirely on the extraction agent’s validation layer; without it, errors would have propagated into bid submissions before anyone noticed.
AgentOps is a different discipline from service monitoring. Both deployments required a monitoring approach specific to agent failure modes. Unlike traditional services that fail visibly, agents can degrade subtly — outputs that pass technical validation but drift off-brand, responses that are technically correct but declining in quality over time. Monitoring needs to track business metrics (bid accuracy, content engagement rates) not just system metrics (latency, error rates).
When Autonomous Agents Deliver ROI
The two cases above succeeded because the problems they addressed shared a specific profile: high-frequency processes, document or data-heavy workflows, and scenarios where human latency created quantifiable financial loss. The travel agency’s overnight revenue gap was a direct consequence of human bandwidth. The tender team’s strategic capacity was blocked by administrative volume.
Gartner research projects that 40% of enterprise applications will embed task-specific AI agents by end of 2026. But the same research body warns that over 40% of agentic AI projects will be cancelled by end of 2027 — primarily due to unmet ROI expectations and governance failures.
The difference between the successful 60% and the cancelled 40% is not access to better technology. It is a clearer answer to three questions before build starts: What specific process bottleneck does this agent clear? What does success look like in measurable terms at six months? Who owns the system after it is in production?
Automating high-frequency, rule-heavy, document-intensive workflows where skilled people are spending time they should not have to spend is where the ROI case is strongest and most consistent. Broad automation of judgment-intensive workflows — where the value of the human is the judgment itself — carries a very different risk profile and a lower probability of delivering the numbers.
How we approach this at Insoftex
Both builds in this article share a quality that we now treat as a selection criterion rather than a happy accident: the problem they addressed had a measurable unit of loss attributable to a specific process bottleneck. The travel agency’s overnight revenue gap was countable in unsold seats. The tender team’s administrative overhead was measurable in senior hours spent on work that did not require senior judgment. That measurability is what makes the ROI case defensible — and what makes it possible to scope the build correctly before committing to it.
Before either production deployment, we ran a scoping engagement that answered the key question before code was written: does the workflow structure, data access, and governance context support an autonomous agent deployment? In the tender case, the critical unknown was whether the tender portals could be scraped reliably enough to feed the parsing pipeline. In the travel agency case, it was whether the tour management system could be accessed by the agent infrastructure without a manual API build. Both questions were answerable in the scoping phase. Either answer could have changed the build.
The monitoring architecture for both systems required design decisions that most teams make too late: what constitutes degraded agent performance, who owns the alert, and what does remediation look like? An agent that drifts off-brand or loses extraction accuracy does not fail visibly — it produces progressively worse outputs. Defining the performance baselines and alert thresholds before the system goes live is what makes those drifts diagnosable rather than mysterious.
Evaluating an agentic AI build for a specific process bottleneck? Our Product Pilot maps the workflow, scopes the architecture, and delivers effort estimates before you commit to a build. Fixed scope, three weeks, senior engineers from day one.
Frequently Asked Questions
What types of business processes are best suited to autonomous AI agents?
Autonomous agents deliver the most consistent ROI in three process profiles: high-frequency workflows where volume creates a human bandwidth constraint, document and data-heavy processes where most of the work is extraction, classification, and routing rather than judgment, and time-sensitive operations where human latency produces quantifiable financial loss. Both case studies above fit this profile: the tender build's 400% submission increase came from eliminating the volume bottleneck; the travel agency's revenue recovery came from eliminating the overnight latency gap. Processes where the primary value is human judgment — complex negotiations, regulatory interpretation, strategic pricing decisions — are not candidates for full automation, though they may benefit from agents that handle surrounding administrative work.
What is AgentOps and why does it differ from standard service monitoring?
AgentOps is a monitoring discipline specific to the failure modes of AI agent systems. Traditional service monitoring watches for latency spikes, error rates, and downtime — failures that are visible and immediate. Agent failures are frequently subtler: outputs that pass technical validation but drift off-brand over time, predictions that are technically correct but declining in quality as data distributions shift, or behaviour that changed after a model API update without triggering any error. AgentOps monitoring tracks business-level metrics — bid accuracy rates, content engagement, decision quality — not just system-level metrics. It requires someone with the authority to act when those metrics fall below threshold, not just the responsibility to notice.
How should we choose between LangChain, LangGraph, and other orchestration frameworks for an agentic build?
Framework selection depends on the structural requirements of the workflow. Sequential pipelines with well-defined steps and limited inter-agent state dependencies are well-served by LangChain or PydanticAI. Systems with branching logic, multiple specialist agents sharing state, or human-in-the-loop approval gates at decision points require LangGraph's graph-based orchestration and LangSmith's audit trail. The tender build used LangChain for a sequential pipeline; the travel agency build required LangGraph because the compliance guard's revision routing created branching state that a linear framework could not manage cleanly. We cover the full framework comparison in our article on AI agent frameworks in 2026.
What does it cost to build a production-grade autonomous agent system?
The range is wide depending on the number of agents, integration complexity, governance requirements, and whether existing data infrastructure is AI-ready. A focused two-to-three agent system addressing a well-defined workflow bottleneck with accessible data can reach a production MVP in 8–14 weeks. Multi-agent systems with real-time data integrations, compliance audit trails, and human-in-the-loop gates typically run 18–28 weeks for a production MVP. Infrastructure and tooling costs at production scale — LLM API costs, vector database, orchestration compute, observability tooling — need to be scoped against the workflow economics: what is the per-unit cost of the process being automated, and what does that cost at the target volume?
What governance elements are non-negotiable for a production agent deployment?
Four elements are required before any production agent deployment: read-only integrations to authoritative data sources (agents should read from systems of record, not write without validation), schema validation on all agent outputs (errors caught at the boundary, not propagated downstream), comprehensive audit trails (what each agent received, reasoned over, and produced — essential for debugging and for regulated industries), and a named post-deployment owner with monitoring access and the authority to act when performance metrics fall below threshold. Human-in-the-loop approval gates for high-impact decisions are strongly recommended during initial deployment and can be relaxed once the system demonstrates consistent performance. All four elements cost more to retrofit after deployment than to design in from the start.