The Redesign Dilemma. Why Automating Legacy Workflows Bleeds Capital

Author: Indira Hayams. Director at Via Motus.

This is part 2 of our previous article about AI economics or Tokenomics. In this post also available in LinkedIn, I offer some reflections on improvements that can be made to reach the stage of tangible benefits in the P&L and successfully deploying AI. As all about AI, concepts may have evolved or their weights have changed. But it is a topic that is more than worthy to follow up these days.

In Part 1, we discussed the financial failure modes breaking enterprise AI budgets: the stark divide between flat seat licenses and Usage-Based Billing (UBB), autonomous agent loops running up exponential bills, and the costly reality of Context Window Creep, were highlights in our first article about the topic.

The data was clear: while up to 89% of organizations report generative AI experimentation, only 6%, the true AI high performers, attribute more than 5% EBIT impact to its deployment.

When finance leaders see rising token invoices without corresponding bottom-line gains, the immediate reaction is often to blame the vendor, negotiate deeper API volume discounts, or order the HR or IT team to mandate another round of "Prompt Engineering" workshops.

That instinct diagnoses the symptom while completely missing the disease.

One critical topic left out of the main discussion in Part 1 is the workforce skilling paradox. Just a few years ago, enterprise adoption meant training employees on "Prompt Engineering" to drive usage. Current AI economics demand a fundamentally different approach: developing workforce AI Fluency and shifting toward Context Engineering, an operational discipline focused on workflow redesign rather than finding magic words, calibrated to organizational maturity.

If we want AI to deliver real returns on the P&L, we must confront an uncomfortable operational truth: layering autonomous systems on top of unoptimized, legacy manual processes simply speeds up organizational dysfunction at $50 per million tokens.

A Lesson from the Field: Selling to the Enterprise

During my years negotiating enterprise agreements, understanding objections to latest updates and upgrades and supporting sellers positioning everything from Productivity to Business Applications to Cloud Infrastructure at Microsoft, we saw a nearly identical pattern play out during early SaaS rollouts.

Whenever an organization took an inefficient, paper-heavy approval chain and simply digitized it field-for-field into a modern cloud platform, the deployment invariably stalled, costs soared, and users rebelled. The software was blamed, but the root cause was obvious: they hadn't transformed how work was done; they had merely electrified an existing bottleneck, and what about the state of data? This is material for another article. Let’s not lose focus on this one.

Generative AI repeats this exact mistake; only this time, the meter runs on every single turn of inference.

The "Layering" Trap: Why Legacy Processes Break Models

In traditional enterprise software, a clunky process was forgiven because human labor provided the connective tissue. If a spreadsheet had missing fields or a policy document was ambiguous, an employee used judgment, walked across the hall, and resolved it.

When you inject generative models or multi-turn agents into that exact same fragmented workflow, three breakdowns occur immediately:

  1. Context Bloat & Attention Rot: Because the legacy workflow relies on fragmented data repositories, the system prompt must ingest massive grounding documents, extensive chat histories, and sprawling tool schemas just to orient the model. As Anthropic engineers observed, LLMs operate under a finite "attention budget". As input tokens multiply, models suffer from context rot: a measurable degradation in retrieval accuracy, reasoning fidelity, and system latency.

  2. Unconstrained Autonomous Drift: Legacy workflows rarely have mathematically bounded handoffs. When an agent encounters an edge case without strict stopping logic, it enters a recursive loop, re-querying vector databases and executing redundant tool calls before returning an output. As Shruti Koparkar pointed out regarding NVIDIA's tokenomics research, agentic turns compound non-linearly; an unbounded loop burns through thousands of billable operations in seconds without delivering verified output.

  3. The Entry-Level Pipeline Breakdown: Recent empirical labor research across millions of workers from Stanford economists using ADP payroll microdata, reveals that AI exposure is already reshaping headcount. Not through base salary cuts, but via a documented slowdown in entry-level hiring in highly exposed fields. Historically, young professionals handled the codified, checkable steps while building the tacit institutional knowledge needed for senior roles. If an enterprise abruptly strips away these entry-level tasks to automated agents without redesigning the end-to-end workflow, it will create an unmonitored execution void where no one is left to audit the agent's work.

You cannot prompt your way out of a broken operating model. True enterprise scale requires evolving from Prompt Phrasing to Context Architecture, and ultimately to Process Re-engineering.

The C-Suite Discussion: 4 Foundational Questions for CXOs

Before authorizing the next AI budget expansion or enterprise platform license, C-suite leaders need to align on architectural boundaries. Here are four questions every executive committee should debate:

1. For the Chief Information Officer (CIO):

"Are we treating AI as an isolated tool bolted onto existing software, or are we redesigning the runtime context pipeline?"

  • Reality: A user’s prompt is barely 5% of the payload processed by the model; the remaining 95% is architectural context: system instructions, RAG retrieval chunks, memory states, and tool schemas. If our engineering team isn’t actively practicing Context Engineering (implementing prompt caching, state compaction, tool result clearing, and dynamic routing), we are paying a 90% structural tax on every inference.

2. For the Chief Operating Officer (COO):

"Where does our process differentiate between 'Delegated Execution' and 'Supervised Decisions'?"

  • Reality: The goal of workflow redesign is not 100% blind autonomy; it is bounded acceleration. Automative tasks (where the model executes tasks with minimal intervention) must have deterministic validation gates, while augmentative tasks (collaborative analysis and complex problem-solving) must empower experienced personnel rather than removing them from the loop. Where are our circuit breakers when an agent hits an ambiguous exception?

3. For the Chief Human Resources Officer (CHRO):

"Are we upskilling the workforce for 'magic prompt words,' or are we building organizational AI Fluency?"

  • Reality: The era of teaching prompt tricks is over. The enterprise skill gap today is problem decomposition: training business analysts and domain experts to deconstruct workflows into discrete, structured, and auditable tasks that agents can reliably execute without budget-busting context creep.

4. For the Chief Financial Officer (CFO):

"Are we still evaluating AI success through vanity adoption metrics, or by Accepted Work per Dollar (AWpD)?"

  • Reality: Tracking daily active users or total prompts run incentivizes waste. High performers measure unit costs: What is the cost per verified compliance audit? What is the cost per resolved customer transaction? How much acceptable compliant work was generated per dollar spent on compute and tokens?

The Organizational Maturity Matrix: Where Does Your Enterprise Sit?

Workflow redesign is not a one-time project; it reflects organizational maturity and financial discipline. Unfortunately, the format of a table does not fit in the blog. In the LinkedIn version, I included a table that connects Maturity Levels (Experimental, Managed and Strategic) with four categories: Workflow integration. Workforce Capability. Context & Architecture and FinOps Metric.

 
The Executive Takeaway

AI value does not accrue to the companies that spend the most on compute; it accrues to the organizations that exhibit the greatest architectural and operational discipline.

If your AI initiatives are struggling to demonstrate financial impact, resist the urge to buy more seats or run more prompt seminars. Take a step back, look at the end-to-end workflow, strip away the legacy friction, and architect the operational context before letting the model execute.

Only then does AI stop being an unpredictable variable expense on your balance sheet and become the structural engine of EBIT growth it was promised to be.


References

Next
Next

The budget pieces for AI Adoption