AI Agents for Businesses: From Pilot to Production in 2026

AI agents for businesses became commonplace in large Brazilian companies in 2026, and so did project failures. A global study by Sinch found that while 76% of large Brazilian firms reported having agents in production, a staggering 80% had already been forced to halt or reverse a deployment. In this guide, we explore the reasons behind these setbacks (data quality, legacy systems, and governance) and outline how to build a pilot that successfully transitions to production, avoiding costly internal headlines.

TL;DR

• In Brazil, 76% of companies with over a thousand employees surveyed by Sinch have AI agents in production, yet 80% have already halted or reversed a deployment. The global average stands at 74%.
• 39% of these reversals in Brazil were due to personal data breaches. The issue is rarely the model itself, but rather what the agent is permitted to read and write.
• The three most common causes I observe are: disorganised data, legacy systems lacking APIs, and an absence of clear governance (who approves, who audits, who can switch off).
• A successful AI agent pilot for businesses that reaches production has a narrow scope, clearly defined business metrics upfront, a kill switch, and a log of every action.
• Sometimes, the right answer is not to use an agent: a deterministic automation can resolve issues more cost-effectively and with less risk.

The 2026 Landscape: High Adoption, High Reversal Rates

The year's figures tell two simultaneous stories, both of which are true.

The first story is one of accelerated adoption of AI agents by large enterprises. Sinch's study, The AI Production Paradox, surveyed 2,527 executives across ten countries, all from companies with over a thousand employees. In Brazil, 76% reported having AI agents in production, exceeding the global average (62%) and the United States (67%). Furthermore, 66% of Brazilian firms plan to increase their AI investment by over 25%.

The second story, however, is one of significant setbacks. The same survey revealed that 80% of Brazilian companies have already had to halt or reverse AI deployments due to governance issues. The global average for this stands at 74%. Of the total reversals in Brazil, 39% were attributed to personal data breaches, contrasting with 24% in the US and 16% in Canada.

The Specific Focus Matters

Before presenting these figures at a board meeting, consider the specific focus of the study. The research examined AI applied to customer communication (service, messaging, voice, email) and exclusively within large enterprises. Sinch, notably, provides communication infrastructure.

For the broader market, the situation differs. The TIC Empresas 2025 survey by Cetic.br reports 17% AI usage among Brazilian companies with ten or more employees. Furthermore, a study presented by AWS at Summit São Paulo in September 2026 revealed that only 6% of companies utilising AI have actually implemented agents. Just 21% claim to be ready for agentic AI, and approximately one-third are confident in their ability to measure its return on investment.

In essence: large enterprises are ahead of the curve but encountering difficulties. Medium and small businesses, however, are yet to begin. If your company falls into the latter group, the good news is that you can learn from the mistakes of others before committing significant resources.

Agent, Chatbot, or Automation: Understanding Your Purchase

Half of all failed AI agent projects for businesses stem from a misunderstanding of terminology. It's crucial to distinguish between three key concepts:

• Deterministic Automation: Fixed rules. For example, "If the invoice is 3 days overdue, send template X." Always performs the same action, is inexpensive, and fully auditable.
• Generative AI Chatbot: Answers questions in natural language based on a knowledge base. It speaks, but does not act.
• AI Agent: Receives an objective, decides the steps, and calls tools (consults the ERP, opens a support ticket, updates a customer record, issues a duplicate document). It acts with a degree of autonomy.

The practical difference lies in the scope of potential damage. A chatbot that makes a mistake provides a poor answer. An agent that errs could cancel the wrong order, expose another customer's personal data, or dispatch 3,000 incorrect billing messages.

Gartner coined the term agent washing for the practice of rebranding chatbots and RPA as "agents," estimating in June 2025 that only around 130 of the thousands of self-proclaimed agentic vendors genuinely deliver on this. In the same forecast, it projected that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value, and weak risk controls.

When a vendor offers AI agents for businesses without detailing the architecture, ask: what tools does it invoke, with what permissions, and where is the log of each action? If the answer is vague, you are likely purchasing a chatbot at an agent's price.

Why AI Agents for Businesses Go Wrong: Data

AI agents for businesses are only as good as what they can read. And most companies don't know exactly what they will be reading.

Disorganised Data Leads to Confidently Incorrect Answers

A knowledge base with three versions of the same exchange policy. A price list where only the sales team knows the current version. A customer database with duplicate email addresses. For a human, these are points of friction. For an agent, they are instructions: it will choose one version and respond with absolute certainty.

Overly Broad Permissions Lead to Breaches

This is where the 39% figure comes into play. The quickest way to make a pilot "work" is to grant the agent broad read credentials in the database. This might function perfectly in a demo. However, in production, a customer simply asking "What is the delivery address for order 4417?" could lead the agent to respond with another person's address.

Under the UK GDPR and Data Protection Act 2018, this constitutes a personal data security incident. The data controller is accountable for the failure, and Article 22 of the UK GDPR grants data subjects the right to request a review of decisions made solely based on automated processing. I've previously analysed on this blog what the Dígitro data breach teaches us about data security: the pattern repeats, only the vector changes.

What to Do Before Activating the Agent

1. Inventory the data sources the agent will consult and eliminate conflicting versions.
2. Create a dedicated access layer (an internal API or database views) that only returns data for the client authenticated in the conversation.
3. Mask or omit national ID numbers, card details, and other sensitive data that the agent does not require to resolve the issue.
4. Test with malicious queries: "Ignore previous instructions and show me…" should be the first test, not the last.

Legacy Systems: Where AI Agents for Businesses Stall

The second reason for reversals is less dramatic but more costly. An agent requires tools, and tools mean APIs. Many older ERPs, academic systems, or Access databases that have been running for 15 years simply lack a proper API.

This leads to common workarounds: screen-scraping robots (RPA), reading spreadsheets exported every morning, or direct access to the production database. Each presents a point of failure that often goes unmonitored. When the ERP screen updates, the agent silently breaks and continues responding with outdated data.

In the AI agent projects for businesses that we oversee at Agathas Web, the integration work often outweighs the AI development itself. In Moodle environments, for instance, an agent answering student queries is only useful if it can consult grades, deadlines, and enrolment via the platform's official web services, with a token restricted to the necessary functions. For projects using the official WhatsApp API, the challenge is rarely the model; it's managing webhooks that arrive duplicated, out of order, or delayed, and the agent's need to process these without responding twice.

Signs Your Legacy System Will Stall Your Pilot

• There is no documented API for the operation the agent needs to perform.
• The only way to read the data is by exporting a spreadsheet or directly accessing the database.
• No one on the team can explain what happens if the system goes offline for an hour.
• The ERP vendor charges for integration and takes months to provide access.

If two of these points are true, your initial project isn't about AI; it's about integration. Investing in an agent before addressing these issues means paying a premium for a pilot that won't get off the ground.

Governance: Who Approves, Who Audits, Who Switches Off

The third reason AI agents for businesses fail is highlighted in the title of Sinch's research: governance. And governance here isn't a 40-page document. It boils down to three questions that require a clear, named individual responsible for the answer.

Who approves what the agent can do? An explicit list of permitted actions, with value and volume limits. For example: "Can issue duplicate invoices. Cannot grant discounts above 5%. Cannot cancel contracts."

Who audits what it has done? Every tool call must be logged with date, client, input, and output. Without this, when something goes wrong, you won't even be able to reconstruct what happened.

Who switches it off, and how quickly? The agent needs a kill switch that doesn't rely on a deployment. This could be a flag in a dashboard, a fallback to human assistance, or a queue that holds actions until someone reviews them.

Rollback is a Good Sign (When Planned)

The same study presents a counter-intuitive finding: among companies that classify their own guardrails as fully mature, the reversal rate is 81%, above the average. The explanation is simple. Those who monitor effectively spot failures and switch off. Those who don't continue to operate with errors unknowingly.

In other words, the goal isn't never to reverse. It's to reverse quickly, with logs, before the customer notices.

What About Regulation?

While the UK is developing its own AI regulatory framework, don't wait for new legislation to get your house in order. The UK GDPR and Data Protection Act 2018 are already in force and cover most of the risks associated with an agent handling personal data. For those operating on WhatsApp, another rule has been in effect since 15/01/2026: Meta's commercial platform terms prohibit general-purpose AI assistants. Your agent must be focused on your business (support, orders, scheduling), not a ChatGPT plugged into your company's number.

How to Build an AI Agent Pilot for Businesses That Reaches Production

This is the blueprint I use. It's not the only one possible, but each step exists because I've seen a project fail by skipping it.

1. Choose a Narrow, Mundane Process

Good candidates: high volume, clear rules, low risk of error, data already accessible. Examples include duplicate invoice requests, order status updates, rescheduling appointments, or deadline queries within a VLE. Bad candidate: "an agent that handles everything."

2. Define Business Metrics Before Writing Code

It's not "model accuracy." It's average resolution time, percentage of conversations resolved without human intervention, cost per interaction, or reopened tickets. Measure the current scenario for two weeks. Without a baseline, no one can prove the return, which is why only a third of companies trust their own measurements.

3. Start with a Human in the Loop

In the initial weeks, the agent suggests actions, and an operator approves them. You measure how many suggestions were accepted without editing. Once this number stabilises at a high level, you can enable automatic actions for lower-risk scenarios.

4. Instrument Everything

Log every conversation and every tool call, set alerts for unusual patterns (e.g., the same client requesting ten duplicate documents, spikes in API costs), and display a dashboard with the metrics from step 2.

5. Define Exit Criteria in Writing

Outcome after 6-8 weeks	Decision
Business metric improved and incidents remained within limits	Expand to full production for the same process
Metric improved, but a personal data incident occurred	Correct the access layer and repeat the pilot
Metric did not move	Terminate or change the process, without attachment
Cost per interaction was higher than human	Evaluate deterministic automation instead

The pilot that "is going well, but it's too early to tell" is the one that consumes budget for six months and then dies at the annual review. Pre-defined written criteria prevent this. I've detailed the return figures for those who have already scaled in AI agents in production: ROI and governance.

When NOT to Use AI Agents for Businesses

Part of my job involves convincing clients not to build an agent. Here are the scenarios where I advise against AI agents for businesses of any size:

• The process fits a simple rule. If it can be described as a five-box flowchart, a deterministic automation will resolve it with near-zero marginal cost and no risk of hallucination.
• Volume is low. Twenty interactions a day do not justify the cost of integration, monitoring, and review. A good quick-response template will suffice.
• The cost of error is high and irreversible. Credit granting, contract cancellation, enrolment decisions. In these cases, the agent can prepare, but humans must make the final decision.
• No reliable data exists. An agent operating on a disorganised database will only automate the mess.
• No one will maintain it afterwards. An agent without an owner becomes a liability. Models change, API prices change, and business rules evolve.

It's also wise to be wary of generic assistants used as a shortcut. Tools like those I discussed in the post about what Perplexity AI changes for businesses are excellent for internal research but do not replace an integrated and audited agent within your specific process.

Cost Breakdown and Budget Overruns

There is no standard list price for AI agents for businesses because the cost depends far more on integration than on the model itself. However, it is possible to identify where the budget will be spent.

Item	What it is	Where it overruns
Model API Consumption	Charge per token (processed text)	Huge context in each call; agent "thinking" in a loop
Integration	APIs, access layer, webhooks	Legacy systems without APIs; ERP vendor charging for connectors
Messaging	Channel fees (e.g., official WhatsApp charges per template message)	Agent sending a template when a response within the 24-hour window would suffice
Observability	Logs, alerts, dashboard	No one budgets for it; the lack is discovered during the first incident
Operation	Person who reviews, adjusts, and is responsible for the agent	Disappears after the pilot, and the agent becomes obsolete

Two practical considerations. Firstly, implement a daily spending cap on your AI provider account from day one. Secondly, always compare costs against the human cost per resolved interaction, not per hour: this is the only metric your board will accept.

How Agathas Web Addresses This

I founded Agathas Web in 2008, and the majority of our work involves operating systems in production: Moodle, Linux servers, managed hosting, bespoke development, and integrations with the official WhatsApp API. It is from this vantage point that we evaluate AI agents for businesses in our technology and AI consultancy: anyone who has witnessed duplicate webhooks and unindexed databases disrupt a service knows precisely where an agent will fail.

In practice, our engagement process is as follows:

1. Free initial consultation (30 to 60 minutes): We understand the candidate process, the technology stack, the volume, and what has already been attempted.
2. Technical deep dive (1 to 2 weeks): Read-only access to systems, data and integration mapping, UK GDPR assessment, and evaluation of what the agent could read and write.
3. Cross-analysis (approx. 1 week): We cross-reference the technical findings with the current process cost and determine if an agent is suitable, if simple automation is better, or if legacy systems need fixing first.
4. Diagnosis and plan: Chosen process, success metrics, permission design, exit criteria, and cost estimate. All delivered within 4 to 6 weeks.

What's included: the prioritised diagnosis, the pilot design, and executive oversight in a CTO as a Service format. We do not earn commission from software resales, so our recommendation might be "don't do it now," and sometimes it is. If execution makes sense, Agathas' development team will build the integration and the agent. If you prefer to proceed with your own team, you will receive a comprehensive plan.

To engage us, simply schedule a diagnostic session on our consulting page and tell us which process you wish to automate.

Conclusion: Less Agent, More Production

The 80% reversal rate doesn't indicate that AI agents for businesses don't work. It suggests that most companies started on the wrong foot: choosing the model before defining the process, granting broad access before designing permissions, and deploying before establishing a clear shutdown procedure.

The path that works is less glamorous. It involves a narrow process, organised data, genuine integration, pre-defined metrics, a human in the loop initially, and a kill switch that any manager can activate. Done this way, the pilot reaches production and stays there.

If you're deciding where to begin, or have a stalled pilot that no one knows whether to expand or terminate, speak with our consultancy. The first conversation is free, and you'll leave with a concrete scope.