AI Agents in Enterprise Operations: Adoption Plan and Metrics
Forecasts, surveys and measured results are not the same. See what the numbers say about AI agents in enterprise operations and a phased adoption plan with metrics.

Marlos Carmo
Founder of Tolky
Updated on November 15, 2026

Share
There is a growing distance between what large companies say about AI and what they actually do with it. In boardrooms, AI is a mandatory topic in every strategy meeting. In operations, reality tends to be different: an FAQ chatbot, a 90-day pilot that never scaled and a nice PowerPoint presented to the board.
To make this concrete, imagine an illustrative scenario (not customer data): an operation that handles thousands of WhatsApp conversations per month, with a queue of repetitive questions (order status, invoice copies, deadlines). Instead of automating everything, it picks a pilot queue, measures the baseline (volume, response time, resolution and reopening) for 90 days and only expands scope when the metrics support it. That path, not a promise of total transformation, is what tends to work.
The question CTOs and IT Directors should ask is no longer "is AI worth it?", but "which queue, with which limits and which metrics?".
Where to Start: the "Solve, Learn, Expand" Framework
We put the plan before the numbers because it is the most actionable part. The timelines below are a planning reference, not a promise:
Weeks 1 to 2: map and choose. List the processes with the highest volume of repetitive interactions. Calculate the current cost (hours x hourly cost). Choose a process with high volume, low variability and available data. Not the most complex one: the most immediate one.
Weeks 3 to 6: configure and connect. Configure the agent with the process knowledge. Integrate the necessary systems (at minimum, the CRM for context and the ticketing system). Define escalation criteria and test with controlled volume.
Weeks 7 to 12: measure and adjust. Go to production with dense monitoring. Measure autonomous resolution rate, satisfaction and cost per resolved interaction. Identify where the agent errs most and feed it more context and instructions.
From month 4: expand. With the first agent working, the second is faster. The infrastructure is in place and internal trust has been built.
See also the AI maturity test and the guide on AI maturity in companies to find out whether your operation is ready.
Evolution of Enterprise Automation
| Attribute | Traditional Automation (RPA) | Intelligent AI Agents |
|---|---|---|
| Processing | Linear, based on strict rules | Dynamic, based on LLM decision-making |
| Adaptation | Fails if the system interface changes | Adapts to flow changes and looks for solutions |
| Supported data | Only structured data (spreadsheets) | Unstructured data (emails, audio, PDFs) |
| Decision | Human (the system only executes) | Autonomous within configured limits and policies |
An enterprise AI agent differs from a chatbot in three dimensions: memory, reasoning and action. It has context about the customer and the process, plans a sequence of steps toward a goal and executes permitted actions in integrated systems (opening tickets, updating the CRM, querying systems) or escalates to humans when necessary. It is the third wave of automation, after RPA and rule-based chatbots, and the definition is in the post what an AI agent is. The architecture decision and autonomy limits are in agentic AI in business.
The Numbers: Forecast, Expectation and Result
Mixing types of data is the most common mistake on this topic. The table separates each one:
| Number | Type | Source and year |
|---|---|---|
| 40% of enterprise applications with task-specific AI agents by the end of 2026, up from less than 5% in 2025 | Forecast (still to be measured) | Gartner, Aug 2025 |
| 80% of common customer service issues resolved without human intervention by 2029, with a 30% reduction in operating costs | Forecast | Gartner, Mar 2025 |
| More than 40% of agentic AI projects will be canceled by the end of 2027 (cost, unclear value, risks) | Forecast | Gartner, Jun 2025 |
| Average expected return of 171% (192% in the US) | Expectation survey, 1,000 executives | PagerDuty, Mar 2025 |
| 74% of executives report ROI within a year with generative AI | Survey (scope: generative AI, not only agents), 3,466 executives | Google Cloud, Sep 2025 |
| 61% of CFOs say AI agents change how they evaluate the ROI of technology projects | Survey, 261 CFOs | Salesforce |
| Enterprise agentic AI market from US$ 2.6 bn (2024) to US$ 24.5 bn (2030), CAGR of 46.2% | Market estimate (other firms diverge widely: MarketsandMarkets projects US$ 5.26 bn to US$ 52.62 bn) | Grand View Research |
An honest reading: Gartner's forecasts point to strong growth, but the same Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027. That is why the plan above insists on limits and metrics. Google Cloud's 74% refers to generative AI, and PagerDuty's 171% is an expectation, not a measured result. To calculate the return for your case, see the AI automation ROI framework.
Cited Cases (with the Right Context)
- Klarna (company-reported data). In a February 27, 2024 release, Klarna reported that its assistant had 2.3 million conversations in its first month (two thirds of chats), work equivalent to about 700 agents, resolution time down from 11 minutes to under 2, a 25% drop in repeat contacts and satisfaction equal to humans (release). In Q3 2025, the CEO said the agent does the work equivalent to more than 853 full-time agents and that savings were US$ 60 million. This is capacity equivalence, not elimination of positions, and CX Dive also notes that Klarna went back to hiring human agents and that customer service and operations costs rose in the quarter (CX Dive).
- General Mills. According to the press (Food Dive), more than US$ 20 million saved since 2024, with AI models that evaluate about 5,000 shipments a day from factory to warehouse. The "agentic" label comes from specialized media (Food Dive).
- JPMorgan (a 2017 historical case). COiN is a machine learning software that reviews commercial credit agreements. The 360,000 hours were the previous annual work of lawyers and analysts, estimated by the bank, not an audited saving. It is not an AI agent in the modern sense (ABA Journal).
The Three Domains of Enterprise Transformation
1. Customer Service: From Reactive Support to Resolution
Customer service is the most mature use case for agents. A service agent does not just answer questions: it solves problems when it has access to the customer history in the CRM, order status and the ticketing system, and when the actions it can perform are defined by the operation (for example, registering a request and triggering the responsible team). What it cannot solve, it hands to a person with the conversation context.
A practical advantage is consistency: the agent applies the same policy in every interaction, without fatigue. For regulated sectors, this can matter as much as cost efficiency. For the volume and cost scenario, see the guide on customer service with generative AI and how to deflect tickets with AI.
2. Internal Operations: What RPA Could Not Reach
RPA only works well when the process is predictable and structured. Any field out of pattern and the robot breaks. Agents understand variation and natural language and handle exceptions that used to require human intervention. Common cases: document processing (invoices, contracts, orders), internal HR requests in natural language and approval orchestration. For companies, good candidates are email and ticket triage, data reconciliation between systems (CRM, ERP, BI) and B2B lead qualification. The gain of each depends on your process and should be measured in your pilot queue.
3. Data and Operational Intelligence
Agents can also act as an intelligence layer over data that today sits in silos (ERP, CRM, helpdesk, finance): monitoring anomalies and alerting proactively. A customer with falling orders and several open tickets may be at churn risk, and an alert with context to the account manager is a level of proactivity that is hard to reach manually. This is an illustrative example; check what the platform delivers today before planning this case.
Generic Agents vs. Enterprise Agents
The difference is not in the language model, but in the architecture around it. Operations at scale need compliance and audit controls (every action traceable), integration with existing systems, intelligent escalation to humans, configurations isolated by area or customer and availability monitoring. These requirements are invisible in a demo and appear in production, at peak service time. Building everything in-house tends to be slower and requires a dedicated team for maintenance, model updates and monitoring. See also AI governance in customer service.
The Three Most Common Mistakes
Mistake 1: starting with the biggest problem. What works is the opposite: the process with the highest volume, lowest variability and clear success criteria. Winning small first builds trust, data and a technical base.
Mistake 2: treating it as an IT project, not an operations one. Agents need a business owner who defines what "good" is and monitors results. Managed only by IT, they tend to optimize technical metrics (uptime, latency) instead of business ones (resolution, satisfaction, cost per ticket).
Mistake 3: underestimating data quality and escalation to humans. An agent is only as good as the data it has access to, and without an escalation design the result is customer frustration. AI integration with the CRM is the step that guarantees up-to-date data. Also measure with the right metrics: autonomous resolution rate, cost per resolved interaction and recontact rate, not just average handling time.
How Tolky Positions Itself
In the market there are many tools that call themselves "AI agents" but are slightly smarter chatbots. Tolky was built for customer service operations with relevant volume: agents that serve on WhatsApp and webchat, integrate with company systems, hand off to a human agent with the conversation history and give managers visibility into what the agent resolves and where it escalates. To estimate the investment, see how much it costs to implement an AI agent.
AI agents will not transform every company at the same time, and Gartner's cancellation forecast shows that starting without limits and metrics is costly. What separates those who get results from those who only experiment is starting with a pilot queue and expanding based on data.
Want to see this in your operation? Talk to the Tolky team: we map 2 to 3 processes and simulate the flow with an agent in your operation.
Share
Tags
- #ai agents
- #enterprise
- #automation
- #b2b
- implementation
- operations
- cto
About the author

Marlos Carmo
Founder of Tolky
Marlos Carmo is an AI entrepreneur and founder of Tolky, the conversational-era infrastructure and AI CRM that unifies intelligent service, multi-channel support (such as WhatsApp and voice), live CRM, and operational intelligence in a single ecosystem. He is a finalist for the SXSW Innovation Awards and a member of Francesco's Economy, a global network of young entrepreneurs focused on innovation and social impact. He works connecting Artificial Intelligence and digital transformation in projects for large organizations.
Next step
Want to bring this to your operation?
Tati, Tolky's AI assistant, understands your scenario in a few questions, suggests the best path and connects you with someone from the team, with full context.


