Begonia InfoSys All articles
AI & Emerging Technology

The Wiring Behind the Worker: What Agentic AI Costs Before It Saves a Dollar

Begonia InfoSys
The Wiring Behind the Worker: What Agentic AI Costs Before It Saves a Dollar

Photo: Internet Archive Book Images, No restrictions, via Wikimedia Commons

An agent that drafts a support ticket is a demonstration. An agent that closes one, applies a credit, and updates the renewal forecast is an operator — and operators need what demonstrations never do: an identity, a defined scope of authority, a shared vocabulary with every system they touch, and a record of their actions that survives an audit.

The market is selling the second thing. Most enterprises have budgeted for the first. That gap is the integration tax, and it is assessed whether or not the pilot succeeds.

The Demo Stops at the Login Screen

Watch closely during a proof of concept. The agent authenticates as a service account with generous permissions, reads from a replicated copy of production data, and writes into a sandbox. Every constraint that makes production difficult has been temporarily removed. The pilot looks fast because the hard work was done in advance — by someone, somewhere, on a slide that said assumed.

Point that same agent at live CRM, ERP, ticketing, and finance systems, and four questions arrive simultaneously. Who is this agent, technically and legally? What is it permitted to do, and who approved that scope? What does the word customer mean in each of the six systems it reads? And when it acts incorrectly at two in the morning, who finds out, and how?

Few organizations can answer any of these with infrastructure they already own. They answer them with project labor. That labor is the tax.

Four Ledgers Inside the Bill

Identity and delegated authority. Agents require non-human identities with narrow, rotating credentials and a named owner of record. In many US enterprises, service accounts still outnumber employees and are governed by nothing more rigorous than a spreadsheet maintained by someone who left. Adding a fleet of autonomous actors to that estate multiplies an existing weakness into an operational hazard. Delegated authority is harder still: an agent acting on behalf of a sales representative inherits that person's entitlements, and when the person changes roles, the agent does not.

The vocabulary problem. Agents reason over meaning, and enterprises store meaning inconsistently. One system treats the account as the customer; another treats the contact as the customer; a third tracks the billing entity. An agent will reconcile these differences silently and confidently — and be wrong in a way that surfaces only during a quarterly reconciliation. Building a shared semantic layer, the definitions rather than the pipes, is unglamorous work that appears in no vendor statement of work.

Process debt and the exception path. This is where the tax begins to compound. Documented processes describe the happy path; the real organization runs on exceptions handled by people who know which rules can bend. An agent trained on the documented process will execute the letter of a policy that was never the whole policy. It will not know that the three largest accounts skip the standard credit check, or that the Midwest distribution center still runs a manual step because an EDI feed broke during a migration nobody finished. Automating an undocumented exception does not eliminate it. It industrializes it and removes the human who used to notice.

Auditability and reversibility. Regulated functions — financial reporting, credit decisions, claims, collections — demand a defensible record of who decided what and why. Agents produce action logs, not reasoning trails. Every action also needs a reversal path. If an agent posts four hundred invoices against the wrong cost center, can anyone unwind that in hours, or does it become a month-long forensic exercise? Reversibility is a design requirement, not a support ticket.

Doing the Arithmetic First

A business case comparing license cost to projected labor savings is not a business case. It is a price comparison. The honest version carries five lines:

  1. Identity remediation. Count service accounts, orphaned entitlements, and systems without centralized provisioning. Multiply by the effort required to bring each into a governed model.
  2. Semantic layer build. Estimate definition work for the five entities the agent will touch — customer, order, invoice, case, product — across every system of record.
  3. Exception harvesting. Budget interview time with the operators who handle edge cases, then the effort to codify what they actually do.
  4. Observability and rollback. Logging, tracing, and a tested unwind procedure for every write action the agent can take.
  5. Permanent supervision. A named human accountable for each agent, with review capacity funded as an operating cost rather than a launch expense.

Add the lines. If the total exceeds projected benefit in year one, the program is not necessarily wrong — but it is an investment in operational maturity spanning several years, and it should be approved as one.

Why the Shortcut Costs More Later

The tempting alternative is to let agents route around process debt: grant broad credentials, allow heuristic reconciliation of conflicting records, and treat the audit trail as a logging problem for later. That produces a fast pilot and a slow catastrophe. Debt you decline to document does not disappear; it migrates into the agent's behavior, where it becomes invisible, autonomous, and extraordinarily difficult to debug.

Organizations that handle this well sequence differently. They fix identity first, because it gates everything else. They build the semantic layer as a shared asset rather than a per-agent integration. They automate only processes they can already describe, and they keep a human in the loop precisely where documentation is thinnest. Above all, they refuse to deploy an agent whose actions cannot be reversed.

None of this is exotic. It is the ordinary discipline of systems integration applied to a new class of user. Teams treating agentic AI as a wave to ride are discovering that the wave is mostly plumbing. The ones treating it as an integration program — with a funded line for the wiring — are the ones whose digital workers will eventually pay for themselves.

The invoice arrives before the benefit. Plan for that order.

All Articles

Related Articles

Governance in Reverse: Diagnosing the Quiet Coup of Undemocratic IT Decision-Making

Governance in Reverse: Diagnosing the Quiet Coup of Undemocratic IT Decision-Making

The Macros Running Payroll: Inside America's Shadow IT Economy

The Macros Running Payroll: Inside America's Shadow IT Economy

Confirming What You Already Believe: The Hidden Danger of Enterprise AI That Learns From the Past

Confirming What You Already Believe: The Hidden Danger of Enterprise AI That Learns From the Past