The Compounding Dollar Value

How Phase 0 Assessments Quantify AI’s Real Impact

In February 2024, a leading fintech firm and a major AI provider put out a joint announcement that the technology press could not stop repeating. The Swedish fintech’s new AI assistant had, in its first month in production, handled 2.3 million customer conversations – roughly two-thirds of the company’s total chat volume. Average resolution time fell from 11 minutes to under 2 minutes. The assistant was doing, the announcement said, the equivalent of the work of 700 full-time agents and was projected to increase profit by $40 million that year. Its CEO said publicly that AI could already do the work humans do. The firm froze hiring.

It was, by every metric on the dashboard, a triumph.

Eighteen months later, the firm was quietly recruiting human agents again. Customer satisfaction had degraded on exactly the interactions that mattered most – refunds, billing disputes, the emotionally charged edge cases where a confident wrong answer costs more than no answer at all. In May 2025, the firm’s CEO told the press the company had gone too far, and that the relentless focus on cost had produced lower quality. The savings that looked so clear in the announcement had not fully materialized. And the cost of unwinding the decision – recruiting, onboarding, retraining – was a number nobody had put in the original business case.

It is worth being precise about the numbers, because precision is the whole point of this article. The “700 agents” figure was a workload-equivalence calculation, not a layoff count, and the $40 million was a forward projection from a single month’s run rate, not a measured result. That is exactly the kind of arithmetic that gets applauded in an announcement and quietly abandoned in a year-end review.

The firm is not a cautionary tale about bad technology. The AI worked. It did what it was asked to do. What failed was the assessment that came before it: nobody had segmented the work by complexity, nobody had modeled what happens to the third of interactions the bot could not handle well, and nobody had costed the reversal.

This is not a rare story. MIT’s NANDA initiative, in its State of AI in Business 2025 report, examined 300 public AI deployments alongside 150 leadership interviews and found that around 95% of enterprise generative AI pilots delivered no measurable impact on the P&L. The researchers were emphatic that the models were not the problem. The problem was integration into real workflows – what they called the learning gap. Gartner, surveying 321 customer service leaders in late 2025, went further, predicting that by 2027, half of the companies that attributed headcount cuts to AI will have rehired for those same functions under new job titles.

Read those two findings together, and a pattern emerges. The failure is rarely the model. It is the absence of a rigorous, unglamorous piece of work that should have happened before anyone signed a contract.

A Different Playbook

Yet a select group of organizations stands apart—and they are winning. Buried inside the same MIT dataset is the 5% – organizations whose AI initiatives moved the P&L in the first year. When researchers looked at what separated them, the answer was disappointingly mundane. They picked one painful, well-understood process. They understood exactly what it cost them today. They knew what “better” would be worth, in dollars, before they built anything. And they were candid about whether their data, systems, and people were ready to deliver it.

The companies that get this right, in other words, are running a different Playbook. They didn’t start with the technology. They started with a disciplined Phase 0 assessment that quantified the opportunity before a single penny was committed – and then used that assessment to build a roadmap they could defend.

So, what is a Phase 0 assessment? It’s a short, structured diagnostic – typically four to eight weeks – that runs before any AI build, pilot, or vendor selection. It does five things:

• Maps the work as it happens today, not as the process documentation claims it does – volumes, cycle times, handoffs, exception rates, who touches what.
• Quantifies the opportunity in dollars, using conservative, defensible assumptions rather than vendor benchmarks.
• Assesses readiness – whether the data, systems, and workflows can support the use case.
• Prioritizes a small set of initiatives by value, feasibility, and organizational appetite.
• Produces a roadmap with milestones, owners, success metrics, and a stated investment case.

The output is not a strategy deck. It is a decision document that leadership can either fund or decline, with the reasoning visible on the page. The firm’s mistake was not that it lacked a strategy. It was that it never did step one or step three.

Why This Conversation Has Changed

For most of the last decade, AI investment ran on optimism. Boards approved budgets for “transformation” without expecting hard returns in the first year. That era is over. In 2026, the conversation shifted from “Should we invest in AI?” to “Show me the math.” Every department head is being asked to justify spending, every initiative is measured against opportunity cost, and every executive is wary of becoming the next cautionary tale about a multi-million-dollar AI program that delivered a slide deck.

This isn’t a bad thing. It’s a maturing market, and it puts a premium on the discipline of quantifying value upfront – which is exactly what a good Phase 0 assessment does.

How Phase 0 Quantifies AI’s Real Value

Most AI ROI calculations follow a simple model: Saved hours multiplied by hourly cost equals savings. It’s a useful starting point, but it’s also where most business cases lose credibility. Three adjustments make this math defensible.

• Use fully loaded cost, not pay rate. An employee’s true cost to the business includes benefits, overhead, tools, and management time – typically 1.3 to 1.5 times their salary. This is the number that belongs in the calculation.
• Apply a realization rate. Saving an hour of someone’s time doesn’t automatically save the business money. It only does so if that hour is reinvested into higher-value work, used to absorb growth without new hires, or eliminated through structural change. A realistic realization rate is between 40 and 70 percent in the first year, rising as the organization improves at redeploying capacity.
• Subtract the real cost of ownership. Implementation, licensing, integration, change management, and ongoing model maintenance must be deducted from the gross savings to yield the true net figure. And – as the firm learned – a serious business case for any replacement strategy should also carry the cost of unwinding it if it underperforms.

A Phase 0 assessment that does this honestly will produce smaller numbers than a vendor pitch deck. It will also produce numbers that leadership can have confidence in, which is the entire point.

Where the Savings Live: A Department-by-Department View

The most productive places to apply this framework are functions where work is repetitive, data-heavy, and decision-rich. Four stand out.

Finance

• Finance teams spend enormous time on invoice processing, three-way matching, reconciliations, and month-end closing. AI tools now handle a meaningful share of this work end-to-end, with human review reserved for exceptions.
• A mid-sized finance function might process 50,000 invoices a year. If each consumes an average of 12 minutes of human time – 10,000 hours annually – and AI removes 70% of that effort, the team frees up roughly 7,000 hours a year. At a fully loaded cost of $65 per hour, that’s about $455,000 in capacity unlocked, before factoring in faster close cycles, fewer errors, and better cash flow visibility.
• Finance teams often can’t reduce headcount in the short term because the freed capacity gets absorbed by audit demands, regulatory reporting, and analysis that the team never had time for. That’s not a failure – it’s value showing up as quality instead of cost. Phase 0 needs to make this distinction clear so leadership doesn’t expect savings that won’t materialize in a smaller payroll.

Procurement

• Contract review, supplier risk monitoring, spend categorization, and price benchmarking are all processes AI handles well. Manual contract analysis alone can consume 4 to 8 hours per contract for legal and procurement teams combined. For a company handling 500 contracts a year, an AI-assisted review process can cut analysis time by 60 to 80 percent. That’s 1,500 to 3,000 hours back – and more importantly, it’s faster supplier onboarding, earlier risk identification, and better negotiating positions, because the team finally has time to prepare.

The dollar value of better procurement decisions usually dwarfs the dollar value of saved time. A Phase-0 should quantify both the operational savings and the margin impact from improved sourcing.

Demand Planning and Forecasting

• This is where AI’s impact is hardest to measure with the hours-saved model alone, because the bigger value isn’t time – it’s accuracy. A modest improvement in forecast accuracy, even five to ten percentage points, translates directly into lower inventory carrying costs, fewer stockouts, and less write-off of obsolete stock.
• For a business with $200M in inventory, a 10% reduction in carrying cost from better forecasting is $4M to $6M a year, depending on the cost-of-capital assumptions. The planner hours saved are real, but they’re a rounding error next to the working capital impact.

This benefit only materializes if the organization changes how it places orders and manages stock in response to the improved forecasts. AI without operational follow-through is just a more expensive spreadsheet.

Predictive Maintenance

• This one uses different math entirely, which is worth flagging clearly. The savings here don’t come from human hours – they come from avoided downtime, extended asset life, and prevention of catastrophic failures.
• A manufacturing line that experiences four unplanned shutdowns a year, each costing $150,000 in lost production and emergency repair, represents $600,000 in annual exposure. Predictive maintenance programs typically reduce unplanned downtime by 30 to 50 percent, which translates to $180,000 to $300,000 in direct savings – before counting longer equipment life and lower maintenance overtime.

These benefits require sensor infrastructure, data pipelines, and operational discipline that many organizations underestimate. Phase 0 must be honest about the readiness gap.

The Readiness Question: What Actually Kills These Projects

Everything above assumes the AI can access the data it needs, in a form it can use, within a workflow someone will follow. That assumption is where most business cases quietly break. Remember MIT’s conclusion: the failures were not model failures. They were integration failures. A serious Phase 0 spends as much time here as it does on math.

• Data quality. The invoice example above assumes that 50,000 invoices arrive in a consistent, machine-readable format. In practice, they arrive as PDFs, scans, email attachments, and EDI feeds, with supplier names spelled four different ways and line items that don’t reconcile to the PO. Extraction accuracy on clean structured data and on a photographed delivery docket are not remotely the same problem. Phase 0 should sample the real data – not the sample the vendor was given – and measure the exception rate, because the exception rate is what determines whether your realization rate is 70% or 20%.
• An AI model that produces a correct answer on a screen nobody looks at has produced nothing. The value is realized when the output writes back into the ERP, ticketing system, and planning tool. Phase 0 should establish, concretely: does this system have an API? Who owns it? What is the change-request queue for the team that maintains it? Is there a middleware layer, or will each integration be bespoke? These questions determine the cost-of-ownership line in your business case, and an order-of-magnitude estimate routinely underestimates it.
• Workflow and the human loop. This is the one the firm got wrong. Automating 70% of a process means designing what happens to the other 30% – and designing it first. Who handles the exceptions? How do they get context on what the AI has already tried? What is the escalation path, and what is the confidence threshold that triggers it? Where does human review fit, and does it create a new bottleneck that eats away at the savings? If the answer to any of these is “we’ll figure that out in the pilot,” the pilot will figure it out for you, expensively.
• Metrics that survive contact with reality. The firm’s dashboard showed an average. The damage was hidden in the distribution – in the specific quadrant of high-complexity, high-emotion interactions where a wrong answer is worse than no answer. Phase 0 should define success metrics segmented by complexity, not aggregated into a single resolution rate. An overall number that looks healthy can conceal a quality collapse in exactly the interactions that determine whether a customer stays.

None of this is glamorous. All of it is cheaper to discover in week three of an assessment than in month nine of a deployment.

The Compounding Effect: Where the Real Money Is

Here is what most AI business cases miss entirely. The first year’s savings are not the prize; the prize is what happens when those savings are reinvested.

• Suppose a company runs a Phase 0 assessment and identifies $2M in net annual savings across finance, procurement, and demand planning in Year 1. The instinct is to bank that $2M and call the project successful. The disciplined move is to reinvest a portion of it – say half – into the next wave of AI initiatives, expanded data infrastructure, or upskilling the teams now operating with freed capacity.
• By Year 2, the original initiatives are still delivering, and a second wave is coming online. By Year 3, the cumulative impact isn’t $6M ($2M* 3 years). It’s meaningfully higher, because each wave of savings funds the next wave of capability, and because the freed capacity has been deployed into work that drives revenue and margin – not just cost reduction.

This is the compounding dollar. And it only happens when an organization treats AI investment as a capacity flywheel rather than a one-time savings event.

Phase 0 as a Recurring Discipline, Not a One-Time Exercise

This is the part most companies get wrong. A Phase 0 assessment isn’t something you do once at the beginning of an AI program and then forget about. It’s the diagnostic that builds the first roadmap – and then it gets repeated, on a deliberate cadence, to build the next one.

• The first Phase 0 establishes the baseline. It quantifies the opportunity, identifies the highest-ROI starting points, lays out a 12- to 18-month roadmap with clear milestones, and sets the success metrics to be measured.
• The roadmap then gets executed in phases – typically two to four initiatives in the first wave, chosen because they’re high-value, technically feasible, and organizationally ready. Each one has a defined milestone, a measurable outcome, and a clear owner.
• At the end of that cycle, a second Phase 0 assessment runs. But this one isn’t starting from scratch – it’s building on what was learned. It re-evaluates the landscape (because AI capabilities have moved), measures the actual realized savings from wave one (because reality always differs from projections), and identifies the next set of priorities. The realization rates from wave one become the assumptions in wave two. The capacity freed in wave one funds wave two.
• This rhythm – assess, analyze & achieve – is what separates organizations that compound their AI investments from organizations that run isolated pilots forever. Each Phase 0 is more accurate than the last because it’s grounded in real data from the previous wave rather than vendor estimates. Each roadmap is more ambitious than the last because the organization has built the capability to deliver.

A reasonable cadence is every 12 to 18 months. Fast enough to keep up with how technology is evolving, slow enough that each wave has time to deliver real results before the next assessment begins.

What This Means for Leadership

If you’re a business head reading this, the practical implication is straightforward. The companies that will outperform AI over the next three to five years aren’t the ones with the most ambitious roadmaps or the biggest budgets. They’re the ones with the most disciplined Phase 0 practice – the ones who quantify before they commit, measure what they delivered, and use each cycle to build a better next one.

The firm had the best AI available, an enviable partnership, and a CEO willing to move faster than anyone. What it did not have, going in, was an honest answer to a small set of unglamorous questions. That gap costs more than the technology ever saved.

Math isn’t complicated. The discipline is. Phase 0 is how that discipline gets built into the way decisions are made.

That’s the compounding dollar. It doesn’t come from any single AI initiative. It comes from the habit of assessing, executing, and reassessing – turning a one-time investment thesis into a permanent operating capability.

And in a market where every dollar of AI spend is now under scrutiny, that habit is what separates the companies that will lead from those still pitching pilots.

Where to Start

You don’t need a budget cycle or a board mandate to begin. You need one process and four honest questions. Pick the single process in your function that consumes the most hours and generates the most frustration. Then answer:

• What does it cost us today? Not an estimate – volumes, cycle times, exception rates, and fully loaded cost.
• What is “better” worth in dollars? And what realization rate can we honestly defend in year one?
• Are our data and systems ready to support it? Sample real data. Find the API owner. Count the exceptions.
• What happens to the work the AI can’t do? Design that path before you automate the rest.

If you can answer all four with evidence, you have the beginnings of a Phase 0. If you can’t answer them, you have just found out why your last pilot stalled – and that is worth more than another proof of concept.

We run Phase 0 assessments as a fixed-scope, four-to-eight-week engagement that ends in a costed, prioritized roadmap your leadership team can fund or decline on the evidence. If you’d like to talk through what that would look like for your organization – or you want a second opinion on a business case already sitting on your desk, get in touch with YASH. The first conversation costs nothing but a few hours, which is considerably less than the alternative.

• AI Real Impact
• Phase 0
• Phase 0 Assessments