ARTIFICIAL INTELLIGENCE

AI automation looks simple. It isn't.

RPMflex  ·  2026-05-18

Share

Every week a new impressive use case appears. An agent that reads emails and schedules meetings. A system that analyzes contracts in seconds. A tool that generates financial reports without anyone having to build a spreadsheet.

The technology exists. The results are real.

But between the demo and an operation actually running inside a service company, there is a path that most articles about AI do not show. It is a path of mapping, process decisions, business rules, exceptions, integrations and data — long before any line of code or agent configuration.

That is exactly the stretch where most projects stall.

According to Gartner, 40% of AI agent projects will be canceled by 2027. Not because of technology failure — but because of failures in method, structure and preparation.

This post is not about which AI tool to use. It is about what needs to be resolved before any automation makes sense inside a service company.

The types of automation AI can offer

Before talking about failures, it is worth understanding what AI automation can really do. There are four main categories — and each has a different level of complexity and prerequisites.

1. Information-generation automation

AI monitors operational data and produces analyses that used to depend on someone building them manually — or simply did not exist.

For service companies, this translates into:

  • Real margin per project calculated automatically at month-end close
  • An alert when a project consumes 70% of the budgeted hours but is only 40% complete
  • Income statement by customer, contract or cost center with no manual intervention
  • Billing projection based on active contracts and logged hours
  • Identification of customers whose cost to serve exceeds the contracted amount

The manager receives the right information at the right time, without depending on anyone remembering to review the report.

2. Repetitive-task automation

AI performs high-repetition, low-variation tasks — activities that always follow the same logic and today consume the time of people who could be in roles that require judgment.

  • Issuing NFS-e (electronic service invoices) when a billing trigger is reached
  • Sending a collection notice when a payment slip is overdue
  • Classifying and routing support tickets by type and urgency
  • Allocating financial entries by cost center based on predefined rules
  • Logging hours from activities completed in projects

3. Communication and service automation

AI interacts with customers and the team at specific points of the operation, without depending on a person being available at that moment.

  • Answering frequently asked questions on the support portal, at any hour
  • Automatically updating the customer on project status
  • Initial screening of new contacts before they reach the sales team
  • Automatic reminders to the team about deadlines or pending timesheets

4. Assisted-decision automation

AI does not decide — it organizes the context so the manager can decide better and faster.

  • Before renewing a contract, consolidating that customer's history of hours, margin, satisfaction and deviations
  • When opening a proposal, bringing references from similar projects with real cost and outcome
  • Flagging contracts at risk of churn based on behavior patterns
  • Identifying professionals with available capacity for allocation before the manager asks

Why projects fail — the honest version

The data is consistent and comes from independent sources:

  • Gartner: organizations without proper data practices for AI will abandon 60% of their projects through 2026
  • Gartner: 63% of organizations do not have — or do not know whether they have — the data management practices needed to support AI
  • Fivetran: 42% of companies reported that more than half of their AI projects were delayed, underperformed or failed because of data readiness problems
  • McKinsey: almost two-thirds of companies have experimented with AI agents, but fewer than 10% have scaled them to deliver tangible value — and eight in ten cite data limitations as the main obstacle
  • McKinsey: only 10% to 20% of isolated AI experiments from the past two years scaled to produce measurable business results

The problem is not the technology. It is how the project is conceived and prepared.

Four causes repeat most often:

Cause 1 — The company underestimates the complexity of the process

This is the starting point of many projects that fail.

The idea seems straightforward: "I want AI to read an expense document — an invoice, a statement, a receipt — and post the financial entry automatically, allocating the amount among the projects involved."

Simple to describe. Complex to implement.

Before the automation works reliably, someone needs to have answered — and documented — each of these questions:

About the document

  • Did the document arrive as a PDF, XML or image? Does the format change how the AI reads it?
  • Are the fields needed for the entry legible and standardized?
  • Is it an electronic invoice with structured XML or a document with no defined format?

About the supplier

  • Is the supplier registered in the system? If not, does the AI suspend the entry, create the record automatically or alert someone responsible?
  • If registered, does the data match the invoice — tax ID, legal name?

About the allocation

  • Which projects receive the allocation? Is that information on the invoice or does it need to be found in another system?
  • Is the allocation by fixed percentage, by hours consumed in the period or by manual criteria?
  • What if one of the projects is closed? Where does the amount go?

About the entry

  • Is the entry posted as a provision or as actual?
  • Which cost center, ledger account and expense category receive the amount?
  • Is there an amount limit above which the entry needs approval before being posted?

About exceptions

  • What if the invoice amount differs from the amount expected in the contract?
  • What does the AI do when it cannot determine the correct answer — stop, assume a default or escalate to a human?

Every question without a documented answer is a point where the automation will stall, fail silently or generate an entry someone will have to correct manually later.

That is exactly where the manager loses confidence in the result. And the project dies.

Cause 2 — The data is not ready

AI agents do not invent information. They process what exists.

If data is fragmented across systems that do not talk to each other — hours in a spreadsheet, expenses in another system, billing in a third — the AI will work with partial slices of reality.

The result: imprecise analyses, alerts that fire when they should not, incorrect allocations, reports that do not match.

AI does not solve data fragmentation — it executes it faster. Automating on top of an inconsistent base does not eliminate error. It only makes it harder to trace.

According to a Fivetran survey, 68% of organizations with less than half of their data centralized reported revenue losses from AI projects that failed or were delayed.

Cause 3 — The specification was skipped before starting

The example in Cause 1 shows what happens when the questions have not been answered. This cause explains why they are often not even asked.

Specifying an automation means documenting, before any development:

The main flow

The path that happens when everything is right: who triggers the process, what data comes in, which systems are consulted, what is produced as a result and where it goes.

All known exceptions

What happens in each non-standard situation. For each deviation: does the AI execute, suspend, alert or escalate for human approval? This mapping needs to be exhaustive.

The decision rules

Who decides what, based on which criteria and within what timeframe. Without this recorded, the AI has no parameter to act consistently.

The limits of the automation

Where the AI runs autonomously and where it must stop and wait for a human decision. This limit is about governance — not technology.

This work is not software development. It is process and governance work. And it is exactly what companies tend to skip because it seems bureaucratic — until the project stalls in production.

Cause 4 — The same project does not work for every company

This point rarely appears in discussions about AI — but it is decisive.

Two seemingly similar consulting firms may have completely different contract models, distinct criteria for measuring project progress and billing methods that look nothing alike.

An agent configured for company A, with its rules, will not deliver the right result for company B without deep reconfiguration.

Taking another company's automation project and trying to replicate it without your own mapping is the fastest way to automate the wrong thing — looking right, which is worse.

What needs to be resolved before starting

Regardless of which tool or vendor is chosen, there is a set of questions every service company needs to answer before starting any AI automation project:

About the data

  • Does the data the automation will need exist and is it recorded?
  • Is it accessible in a single system or scattered across different sources?
  • Is it reliable — or does it depend on someone consolidating it manually?

About the process

  • Is the process to be automated documented?
  • Have the known exceptions been mapped?
  • Is there a clear rule for each deviation from the flow?

About governance

  • Who validates whether the AI's result is correct?
  • Where does the automation run on its own and where does it need human approval?
  • What happens when the AI encounters a case that was not anticipated?

About the goal

  • What exactly will this automation deliver?
  • How will we measure whether it worked?
  • What is the current cost of this process done manually?

Without these answers, the project starts in the dark — regardless of which technology is used.

The role of data integration in this process

Service companies have a particular challenge: their most important data is invisible.

Hours worked. Scope consumed. Activities performed. Service requests handled.

This data exists — but is rarely organized in a single place. And when it is not integrated, no automation can work with it reliably.

  • The automation that calculates margin by project needs hours, expenses and billing to be in the same environment.
  • The scope-deviation alert needs the planned and actual budget to be connected.
  • Automatic allocation of entries needs cost centers, projects and suppliers to share the same data structure.

Integration is not a technical detail. It is the operational prerequisite of any automation that involves financial results.

Conclusion

AI automation offers real gains for service companies — in visibility, speed and quality of decisions.

But those gains depend on a foundation that most companies do not yet have in place: integrated data, documented processes, mapped business rules and anticipated exceptions.

Without that foundation, the AI project will work in the pilot — and die in production. Or it will automate the wrong processes much more efficiently.

The question worth asking before any initiative is not "which AI tool will we use".
It is: "what will the automation need to know to work — and is that knowledge recorded anywhere?"

RPMflex offers AI automation integrated with the ERP for service companies — as custom projects, developed from the mapping of each operation's specific rules and needs. The platform already delivers the integrated data foundation any automation will need: projects, hours, expenses, billing and finance in a single environment.

Get to know RPMflex →

Sources and references

All data cited in this article was verified at the original sources before publication.

Gartner

  • Over 40% of agentic AI projects will be canceled by 2027 — Gartner Newsroom (Jun. 2025)
  • 60% of AI projects will be abandoned due to lack of AI-ready data — Gartner Newsroom (Feb. 2025)
  • 63% of organizations do not have adequate data management practices for AI — same survey as above (Q3 2024, 248 data leaders)

Fivetran

  • 42% of companies reported failure or delay in more than half of their AI projects due to data readiness problems — Business Wire (May 2025) · Full report
  • 68% of organizations with less than half of their data centralized reported revenue losses — same survey as above

McKinsey & Company

Share this article

Take the next step

Fewer spreadsheets. More control. More profitability.

RPMflex integrates sales, projects, time tracking and finance on a single platform — from sale to billing.

Try it free

Start now, no commitment.

More posts

Back to articles →
Tax Management

Receita Federal's Sintonia A+ Seal: what it is, the criteria and how your company earns it

Receita Federal's Sintonia A+ Seal: learn what it is, the compliance criteria and the step-by-step guide for your company to ach……

Tax Management

NFS-e through the National Issuer becomes mandatory for Simples Nacional companies in September

Starting September 1, 2026, micro and small businesses under the Simples Nacional regime are required to……

Financial Control

Sales Commission: how to avoid mistakes, motivate the team and protect cash flow

Learn how sales commission management affects motivation, cash flow and controllership in service companies. L……