Most enterprise AI strategies start with the same question:
How do we get more employees to use AI?
Companies buy thousands of licenses, run prompt-training sessions, and track active users, prompts, and tokens.
The dashboard goes up. The work often stays the same.
This is Goodhart’s law in action: when a measure becomes a target, it stops being a good measure. AI adoption was supposed to indicate progress. Now companies optimize for visible activity rather than better operations.
Giving every employee a separate chatbot rarely fixes this problem. Instead, companies should run a central agent in the background. That agent connects directly to company systems and brings employees in only when they must judge or approve an action.
Adoption is high. Transformation is not.
McKinsey’s 2025 State of AI report found that 88% of surveyed organizations used AI in at least one business function.
Only about one-third of respondents scaled their AI programs, and only 6% generated significant value or at least 5% EBIT impact.
Deloitte found a similar divide in its 2026 enterprise AI report:
- 37% used AI on surface tasks without changing underlying processes.
- 30% redesigned core processes around AI.
- 34% pursued deeper operational changes.
A company can report widespread adoption while its core workflows remain untouched.
An employee uses a chatbot to rewrite an email. A team launches a pilot. A department creates its own collection of prompts and assistants. All of this counts as adoption.
Proving real operational progress requires tangible results: teams complete more work, make fewer errors, resolve customer requests faster, and spend less money.
Personal AI creates a predictable barbell
In our client work, employee AI usage repeatedly follows the same distribution.
Roughly 10% of employees consume 90% of the tokens: they build advanced workflows and use AI all day.
Another 20% use it occasionally and get some value.
The remaining 70% barely use it at all.
Training helps, but the distribution persists. Prompting models effectively requires deep context, thorough review, and a clear understanding of model limits.
A few employees master these skills. Most employees use AI only when a clean interface makes the benefit immediate.
When every employee uses a separate AI tool, operations fragment:
- different instructions,
- different context,
- different tools,
- different quality standards,
- different approval rules,
- and no consistent record of how work was completed.
The company becomes dependent on individual prompting skills, and valuable workflows stay trapped in private chats.
Put the agent at the center
A better architecture puts one governed agent system at the center of the organization.
This central agent connects directly to existing work systems. It maintains business context, enforces shared policies, coordinates repeatable workflows, and logs every action.
Employees interact with the system according to their roles: they supply missing context, review exceptions, correct drafts, and authorize high-stakes actions.
Work begins the moment an event happens: an incoming invoice, a customer ticket, an account update, or a reporting deadline triggers the workflow automatically. Teams no longer depend on individual employees remembering the right prompt.
AI becomes shared operational infrastructure.
Give each part of the workflow to the right system
A central architecture divides work cleanly: deterministic steps run in software, bounded judgment goes to agents, and consequential decisions stay with people.
Deterministic work goes to software
When rules and inputs are fixed, use conventional code.
Frontier models can write this code quickly and cheaply to match how your team works. A model can build an integration, rule, or internal tool once, without running on every live transaction.
Software should validate fields, apply fixed thresholds, match IDs, calculate totals, and sync data across databases. This keeps operations fast, cheap, and predictable.
Bounded judgment goes to the agent
Agents excel when a task requires interpretation, validators can check the output, and mistakes carry low risk.
Examples include:
- classifying incoming requests,
- reconciling inconsistent records,
- preparing reports from several systems,
- drafting responses from account history,
- identifying anomalies,
- and routing unusual cases to the right employee.
A 2026 study of 758 BCG consultants shows why this routing matters. On work within the model’s capabilities, AI users completed 12.2% more tasks and worked about 25% faster.
On a task outside that boundary, AI users were 19 percentage points less likely to reach the correct answer.
Engineers must specify where model judgment helps, how automated checks verify outputs, and when the system escalates cases to people.
Consequential judgment stays with people
People must decide high-risk and irreversible actions. The agent should prepare almost everything leading up to that choice.
Consider a supplier payment. The agent retrieves the invoice, purchase order, and receipt; verifies totals, bank details, duplicates, and policies; flags discrepancies; and drafts the approval record. An employee reviews the nearly finished case, resolves edge cases, and releases the funds.
This division of labor provides a practical target for sensitive work: the agent finishes roughly 90% of the preparation, while a person owns the final judgment and action.
Measure the operation
When an agent runs core operations, work metrics measure real progress.
Track:
- cost per completed case,
- cycle time,
- first-pass accuracy,
- rework rate,
- exception rate,
- human review time,
- customer resolution time,
- and control failures.
Token consumption belongs in the cost calculation.
A meaningful report shows:
The invoice agent completes 61% of cases without intervention. Analysts review 31%, and 8% become exceptions. Average handling time fell from 22 minutes to 9 minutes. Payment release remains under human control. Total cost per completed case fell by 28%, including inference and review.
This report measures operational change, while an adoption chart merely counts software clicks.
The goal is simple: let software execute deterministic rules, let agents exercise bounded judgment, and let people make consequential decisions.
Then measure whether the work got better.